NewMind AI Weekly Chronicles - July'25 - Week IV

NEWMIND AI JOURNAL WEEKLY CHRONICLES
22.7.2025 - .31.7.2025
• The fourth week of July 2025 marked a historic milestone in AI, with major advancements across model development, infrastructure, enterprise adoption,
and research.
• Alibaba released Qwen3-Coder-480B-A35B-Instruct, a 480 billion-parameter open-source model with 35 billion active parameters that rivals Claude
Sonnet 4 in agentic coding and supports a 256K context window extendable to 1 million tokens.
• OpenAI signed a $30 billion annual agreement with Oracle under Project Stargate, securing 4.5 GW of data center capacity—enough to power four million
homes—and expected to triple Oracle’s cloud revenue by FY2028.
• Google introduced LSM-2 for wearable sensor data processing using Adaptive and Inherited Masking, NVIDIA’s NeMo framework enabled reasoning-
capable LLM training in just 48 hours on a single GPU, and Zhipu AI unveiled GLM-4-5-MoE, a 1.8 trillion-parameter model specialized for PowerPoint
generation.
• Enterprise AI deployment surged as Intuit’s agentic solutions saved mid-sized businesses 17–20 hours monthly, IBM exceeded earnings forecasts through
AI-enhanced mainframes, and Freed’s AI scribe scaled to 20,000 clinicians.
• New research pushed boundaries with the Thread Inference Model (TIM) enabling unlimited context length, MegaScience releasing a 1.25 million-instance
dataset for scientific reasoning, and Stanford HAI launching the In Silico Center for computational social science using generative agents.
• Geopolitical AI efforts escalated, with Taiwan unveiling a $510 billion AI strategy, the White House publishing its AI Playbook for global leadership, and
China proposing international AI governance at the Shanghai Cooperation Organization summit.
• Sustainability entered the spotlight as Mistral AI published a full lifecycle analysis revealing that training Mistral Large 2 emitted 20.4 kilotons of CO₂ and
used 281,000 cubic meters of water.
• Safety and security remained a core concern, with Anthropic developing auditing agents for misalignment detection, Amazon launching the NOVA AI
Security Challenge to simulate real-world attacks, and Microsoft introducing the Ladder of Reasoning to assess imaginative reasoning in LLMs.
# Highlights Summary Author Source Date
1.1
Alibaba releases
Qwen3-Coder-480B
On July 22, 2025, Alibaba released Qwen3-Coder-480B-A35B-Instruct,
part of its open-source Qwen3-Coder family. The model is a
By Asif Razzaq 🔗 July 23, 2025

-A35B-Instruct, its
most powerful
open-source
agentic code
model.
480 billion-parameter Mixture-of-Experts architecture with 35 billion active
parameters, supporting native 256 K token context—extendable to 1 million
tokens—and excels at agentic coding, tool use, and browser automation,
matching performance of proprietary models like Claude Sonnet 4.
Developed via large-scale reinforcement learning across 20,000 parallel
environments on Alibaba Cloud, it achieves state-of-the-art results on
benchmarks such as SWE-Bench Verified. Alongside, Alibaba
open-sourced the Qwen Code CLI tool to streamline developer interaction.
1.2
OpenAI secures
massive deal—
$30 billion annually
for Oracle data
center capacity
under Project
Stargate.
OpenAI has confirmed a historic agreement with Oracle to lease 4.5 GW of
data center capacity as part of its expansive Project Stargate initiative.
Valued at approximately $30 billion per year starting in 2028, the deal
supports Oracle’s construction of multiple hyperscale AI data centers
(including the Stargate I campus in Abilene, Texas) and acquisition of tens
of thousands of Nvidia GB200 GPUs. The capacity, equivalent to powering
four million homes, diversifies OpenAI’s cloud infrastructure beyond
Microsoft Azure. Oracle expects the deal to triple its current cloud revenue,
fueling over 50% annual growth by FY 2028.
By Julie Bort 🔗 July 22, 2025
1.3
AWS’s GENIAC
program in Japan
underscores that
scaling foundation
model builds
requires
AWS’s “GENIAC” initiative, part of Japan’s national generative AI
accelerator program, enabled 12 organizations to launch over 127 EC2 P5
and 24 Trn1 (Trainium) instances in a single day to train foundation models
ranging from a 32B multimodal model to a 405B multilingual tourism model.
The initiative proved that large-scale model training is more an
By Keita
Watanabe and
Masaru Isaka
🔗 July 22, 2025

organizational
support, not just
hardware.
organizational challenge than purely hardware-driven—requiring
cross-functional collaboration, reproducible templates, and structured
engagement from AWS Solutions Architects, support teams, and
developers. Because of this, even small teams successfully executed
complex workloads. AWS has already launched cycle 3 after hosting a
hands-on April 2025 technical workshop in Tokyo.
1.4
NVIDIA NeMo
enables training a
reasoning-capable
LLM in
approximately 48
hours on a single
GPU.
NVIDIA’s NeMo framework allows developers to train reasoning-capable
large language models (LLMs) within approximately 48 hours on a single
GPU. Using the open-source Llama Nemotron dataset, which includes over
32 million samples from math, coding, science, and chat domains, NeMo
fine-tunes models to enhance reasoning skills. The training workflow
involves NVIDIA NeMo Curator for data preprocessing and NeMo
framework for efficient model training and evaluation. This method
significantly lowers barriers for researchers and developers, democratizing
access to powerful LLM capabilities and accelerating the creation of
specialized AI models for reasoning-intensive applications.
By Mehran
Maghoumi, et
al.
🔗 July 22, 2025
1.5
Google introduces
LSM-2, a self-
supervised model
adept at learning
from incomplete
wearable sensor
data.
Google's LSM-2 (Large Sensor Model 2) employs Adaptive and Inherited
Masking (AIM) to effectively handle missing data in wearable sensor inputs.
Traditional models often require complete datasets, but AIM allows LSM-2
to learn from fragmented real-world data by masking both artificial and
naturally occurring gaps. Trained on 40 million hours of data from over
60,000 participants, LSM-2 excels in tasks like health condition
By Girish
Narayanswamy
and Maxwell A.
Xu
🔗 July 22, 2025

classification, activity recognition, and continuous health metric prediction.
It outperforms its predecessor, LSM-1, by maintaining higher accuracy
even when significant portions of data are missing. This advancement
enhances the robustness of wearable health technologies in real-world
scenarios.
1.6
Google launches
Gemini 2.5 Flash-
Lite, its fastest and
most cost-effective
model in the
Gemini 2.5 family
Google has released the stable version of Gemini 2.5 Flash-Lite, marking it
as the fastest and most cost-efficient model in the Gemini 2.5 lineup.
Designed to optimize performance per dollar, Flash-Lite offers a balance
between speed and quality, making it suitable for latency-sensitive tasks
like translation and classification. It supports native reasoning capabilities,
which can be toggled on for more complex use cases. Additionally, Flash-
Lite includes features such as a 1 million-token context window, multimodal
input, and native tool support, including Google Search grounding and
function calling.
By Logan
Kilpatrick and
Zach Gleicher
🔗 July 22, 2025
1.7
Alibaba Launches
Its Most Advanced
Open-Source AI
Coding Model to
Date
Alibaba has released its most advanced open-source AI coding model,
designed to rival industry leaders in code generation and developer
assistance. Announced on July 23, 2025, the model supports multiple
programming languages, offers multi-turn coding support, and integrates
debugging capabilities. It is trained on high-quality code repositories and
released under an open license to encourage adoption across global
developer communities. The launch reinforces Alibaba’s growing role in
By Reuters 🔗 July 23, 2025

open-source AI and aims to boost China's competitiveness in foundation
models, particularly in coding-focused applications.
1.8
GPT-5 Expected to
Launch Mid-2026
with Focus on
Accuracy and
Steerability
According to internal updates reported by The Verge, GPT-5 is scheduled
for release in mid-2026, with OpenAI focusing on improving factual
accuracy, steerability, and reliability. Unlike GPT-4o, which emphasized
multimodality and latency, GPT-5 is expected to offer a more stable base
for enterprise and high-stakes applications. The model is undergoing
extensive internal testing, with safety and robustness as core design goals.
This signals OpenAI's pivot toward building AI systems suitable for
regulated industries and long-term deployments.
By The Verge 🔗 July 24, 2025
1.9
Alibaba's Qwen-MT
Sets New Standard
in Multilingual
Machine
Translation
Alibaba’s Qwen team has introduced Qwen-MT, a family of multilingual
foundation models trained for high-quality machine translation across over
100 languages. With parameter sizes from 0.5B to 72B, Qwen-MT models
are open-source and outperform Google Translate, DeepL, and NLLB in
76% of translation directions on FLORES-101. The models leverage a
modular training approach combining pretraining, instruction tuning, and
translation tuning. Qwen-MT also enables zero-shot translation in low-
resource and unseen directions. This release marks a major advance in
open multilingual LLMs and boosts accessibility for global language AI.
By Qwen Team 🔗 July 24, 2025

1.10
NVIDIA Releases
Nemotron-4 340B
and Nemotron-MoE
Models for Custom
AI Agent Training
NVIDIA has unveiled Nemotron-4 340B, a family of large language models
designed to help developers build custom, domain-specific AI agents. The
suite includes a base, instruct, and reward model, plus the open Nemotron-
MoE 122B mixture-of-experts model, which uses only 39B active
parameters per token for high efficiency. These models support NVIDIA’s
NeMo framework and NIM inference microservices, with benchmarks
showing state-of-the-art performance among open models. Developers can
fine-tune using Retrieval-Augmented Generation and Reinforcement
Learning from Human Feedback to boost performance in specific use
cases.
By Udi Karpas 🔗 July 25, 2025
1.11
JAM: A Tiny Flow-
based Song
Generator with
Fine-grained
Controllability and
Aesthetic
Alignment
JAM is a 530M parameter generative model designed for lyric-to-song
synthesis with fine-grained control. Leveraging conditional flow-matching, it
aligns phoneme- and word-level timing to maintain prosodic and phrasing
coherence. A new benchmark, JAME, evaluates generation quality across
musicality and alignment metrics. JAM achieves superior performance over
existing baselines in both automated metrics and human evaluations.
Aesthetic quality is further enhanced using Direct Preference Optimization
(DPO) guided by human preferences. Despite its compact size, JAM
delivers strong controllability and naturalness, establishing a new standard
in lyric-conditioned song generation while maintaining high efficiency and
musical alignment.
By Renhang
Liu, et al. 🔗 July 28, 2025

1.12
Zhipu AI Launches
GLM-4-5 Open-
Source Model
Family with
PowerPoint
Generation
Chinese startup Zhipu AI has released the GLM-4-5 model family, including
the powerful GLM-4-5-MoE, an open-source Mixture-of-Experts model
boasting 1.8 trillion parameters (with 24B active per token). The suite also
includes GLM-4-5-9B and a compact 1.8B version, covering diverse
deployment needs. A standout feature is its ability to automatically generate
PowerPoint presentations from natural language prompts, making it
practical for business and education. These models position Zhipu AI as a
key open-source player competing with GPT-4-level capabilities.
By Carl Franzen 🔗 July 28, 2025
2.1
NVIDIA's NCCL 2.27
introduces
advanced tuning
capabilities to
optimize GPU-to-
GPU
communication for
AI workloads
NVIDIA's Collective Communications Library (NCCL) 2.27 enhances GPU-
to-GPU communication by introducing a dynamic cost model and
scheduler. These improvements enable NCCL to make real-time decisions
on optimal protocols, algorithms, and chunk sizes based on system
topology and message sizes. The library can now run up to 64 Cooperative
Thread Arrays (CTAs) simultaneously, balancing performance and
resource utilization. For platforms with unique configurations, NCCL
supports tuner plugins that allow administrators to override default settings,
By Ben
Williams, et al. 🔗 July 22, 2025

ensuring optimal performance across diverse environments. These
advancements are crucial for scaling AI training and inference efficiently.
2.2
AI-Driven
Mainframe Sales
Help IBM Surpass
Earnings
Expectations
IBM has exceeded Wall Street expectations for Q2 2025, fueled by strong
sales of its AI-enhanced mainframes and hybrid cloud offerings.
Announced July 23, 2025, revenue reached $16.4 billion, boosted by
demand for Power11 systems optimized for enterprise AI inference. The
company’s strategy to embed AI across its hardware portfolio is paying off,
particularly in financial services and government sectors. With AI workloads
increasingly moving on-prem for privacy and latency reasons, IBM’s
approach highlights a resurgence of mainframes as secure, high-
performance AI infrastructure.
By Mike
Wheatley 🔗 July 23, 2025
2.3
Armada Raises
$131M to Deploy
Portable AI Data
Centers in Remote
Locations
Startup Armada has secured $131 million in funding to scale its portable,
self-contained data centers designed for edge computing in remote or
resource-constrained locations. These mobile units house GPUs and AI
accelerators optimized for low-latency workloads, including satellite
communications, defense, and disaster response. Armada’s solution
reduces dependence on cloud infrastructure by bringing compute power
directly to the edge, enabling real-time AI processing where connectivity is
limited. The investment reflects growing demand for decentralized AI
infrastructure beyond traditional data centers.
By Mike
2.4
Intel Scales Back
on Semiconductor
Manufacturing
Amid Strategic
Shift
Intel is dialing down its semiconductor manufacturing investments, delaying
the launch of its $20B Ohio plant and pausing construction on its $25B
Israel fab. The move signals a shift toward capital efficiency amid softening
chip demand and rising competition. CEO Pat Gelsinger reaffirmed Intel’s
IDM 2.0 strategy and AI chip ambitions, including the upcoming Gaudi 3
By Rebecca
Szkutak 🔗, July 25, 2025

launch, but emphasized “disciplined capital” deployment. The strategic
recalibration aims to balance growth with operational agility while
maintaining focus on AI and foundry services. Intel’s pivot aligns with a
broader trend of cautious scaling in the chip sector.
2.5
Huawei Launches
CloudMatrix 384 as
Alternative to
NVIDIA’s AI Stack
Huawei has introduced CloudMatrix 384, a high-density AI server
positioned as a direct alternative to NVIDIA’s AI infrastructure. The system
integrates Huawei’s Ascend 910B AI chips and supports 384 devices per
rack, delivering 2.3 exaflops of FP16 performance—targeting large model
training and inference. CloudMatrix 384 is tightly coupled with Huawei’s
software stack, ModelArts, and CANN, forming a vertically integrated
solution. Amid U.S. export restrictions, Huawei’s launch underscores
China’s push for self-reliant AI infrastructure and challenges Western
dominance in foundational AI compute.
By Mike
2.6
Lisuan G100 GPU
shows promise, at
least in OpenCL —
homegrown
Chinese chip
outguns Arc A770
and RTX 4060 in
new benchmark,
10% slower than
RTX 5060
The Lisuan G100 GPU, a homegrown Chinese graphics processor, has
shown impressive OpenCL performance, outperforming Intel’s Arc A770
and Nvidia’s RTX 4060 in Geekbench benchmarks. With 48 compute units,
12GB of GDDR6 memory, and a 2.0GHz clock speed, it scored over
111,000 points—about 10% slower than the unreleased RTX 5060. This
marks a major leap from its early prototype, which had far fewer resources
and scored just 15,000. Though OpenCL scores don’t directly reflect
gaming performance, the G100 suggests China is making serious progress
in GPU development, narrowing the gap with Western competitors.
By Zhiye Liu 🔗 July 24, 2025
2.7
New AI Chips Aim
to Solve Energy
Chipmakers are developing a new generation of AI processors focused on
dramatically reducing energy consumption, a growing bottleneck in
By WSJ 🔗 July 25, 2025

Efficiency Crisis in
Model Training
training large models. Companies like Cerebras, Tenstorrent, and
EnCharge AI are designing architectures optimized for efficiency over
brute-force performance—targeting lower precision, sparsity, and novel
memory handling. These chips aim to cut energy costs while maintaining
model quality, making AI training more sustainable at scale. With AI
workloads projected to double power demands, these innovations are seen
as critical to enabling continued LLM growth without overwhelming data
center infrastructure.
2.8
GitHub Details
Secure
Architecture for
Scaling Remote
MCP AI Training
GitHub has published a technical guide on building secure and scalable
remote Managed Compute Provider (MCP) servers, enabling developers to
train generative AI models efficiently on third-party hardware. The
architecture leverages Kubernetes for orchestrating compute tasks, a
gRPC-based service mesh for communication, and end-to-end encryption
using SPIFFE identities. This setup supports multi-tenant isolation and
dynamic scaling, critical for large-scale LLM and diffusion model training. It
also addresses trust and security challenges in outsourced AI compute,
paving the way for democratized AI infrastructure access.
By Den
Delimarsky 🔗 July 25, 2025
2.9
SmallThinker: A
Family of Efficient
Large Language
Models Natively
Trained for Local
Deployment
SmallThinker is a series of large language models specifically designed for
efficient local deployment on consumer-grade hardware. Through a
combination of sparse Mixture-of-Experts architectures and optimized
routing strategies, the models achieve high inference speeds and low
memory usage, even on CPUs. A hybrid attention mechanism and
lightweight key-value caching further enhance efficiency. The largest model
(21B) runs at over 20 tokens/second using just 8GB of RAM. Evaluation
across math, reasoning, and translation tasks shows strong performance.
The design emphasizes cost-effectiveness, making advanced language
By Shanghai
Jiao Tong
University
Zenergize AI
🔗 July 28, 2025

capabilities accessible without relying on cloud infrastructure or specialized
hardware.
3.1
Anthropic
Researchers
Identify the “Weird
AI Problem”: When
Thinking Longer
Makes Models
Dumber
Anthropic researchers have uncovered a counterintuitive flaw in LLMs—
dubbed the “Weird AI Problem”—where longer reasoning chains can
degrade performance rather than improve it. The study found that as
Claude models think through problems in more steps, they sometimes
reinforce wrong assumptions or amplify noise, leading to worse answers.
This phenomenon challenges the assumption that more reasoning equals
better reasoning and suggests a need for new techniques to guide or
constrain multi-step thought processes. The findings raise important
questions for agent design, interpretability, and safe model scaling.
By Anthropic
Team 🔗 July 22,
2025
3.2
BEYOND CONTEXT
LIMITS:
SUBCONSCIOUS
THREADS FOR
LONG-HORIZON
REASONING
To overcome context length limitations in LLMs, we introduce the Thread
Inference Model (TIM), a novel architecture that decomposes complex
tasks into hierarchical subtasks. Paired with a runtime system called
TIMRUN, which maintains key-value cache only for relevant reasoning
threads, TIM avoids context overflow, position encoding issues, and GPU
memory bottlenecks. This design significantly improves efficiency and
accuracy on tasks requiring deep reasoning and long-term information
By Hongyin Luo,
et al. 🔗
July 22,
2025

retrieval. With a single model call, TIM + TIMRUN enables virtually unlimited
memory, multi-step tool use, and scalable, structured problem-solving
beyond traditional transformer limits.
3.3
MEGASCIENCE:
PUSHING THE
FRONTIERS OF
POST-TRAINING
DATASETS FOR
SCIENCE
REASONING
MegaScience addresses the gap in scientific reasoning datasets within
open-source LLMs. It introduces TextbookReasoning, a set of 650K
questions extracted from 12K university-level science textbooks across
seven disciplines, paired with verified reference answers. Building on that,
the authors curate MegaScience, a 1.25 million instance dataset compiled
from high-quality scientific sources via ablation studies optimizing data
selection. They also launch a unified evaluation framework spanning 15
benchmarks. Models like Llama 3.1 and Qwen 2.5/3—when trained on
MegaScience—consistently outperform their official instruction-tuned
counterparts, with the largest models showing the greatest gains. All code,
data, and seven trained models are released to support science reasoning
research.
By Run-Ze Fan,
et al.
🔗
July 22,
2025
3.4
GitHub
demonstrates how
Copilot Agent Mode
+ Playwright MCP
server accelerates
UI debugging
On July 22, 2025, GitHub showcased using Copilot Agent Mode with a
Playwright MCP server to automate UI debugging tasks. Developer Chris
Reddington walked through diagnosing layout issues in a Next.js app—like
header overlap and footer spacing—by prompting the agent to test pages,
pinpoint CSS misconfigurations, and propose fixes. The agent leveraged
browser snapshot and interaction tools to inspect elements, apply code
changes, then iterate based on testing results. Emphasizing precise
requirements, the process underscored best practices like updating custom
instruction files, configuring MCP servers in IDE, clearly articulating tasks,
and reviewing iterative outputs to maintain control and accuracy.
By Chris
Reddington 🔗 July 22,
2025

3.5
Can One Domain
Help Others? A
Data-Centric Study
on Multi-Domain
Reasoning via
Reinforcement
Learning
This paper explores how training large language models across multiple
domains—math, code, and logic—affects reasoning abilities. Using
reinforcement learning with verifiable rewards (RLVR), the authors fine-
tune Qwen-2.5-7B models via a new algorithm, GRPO. They find that
certain domains, like math and logic, mutually reinforce learning, while
others, like code, can conflict. Multi-domain training yields more balanced
performance than single-domain fine-tuning. Results also highlight the
importance of reward design, training language, and curriculum strategies.
The study shows how domain interactions influence generalization and
suggests better paths for aligning LLMs to complex, structured tasks
through RL.
By Yu Li, et al. 🔗
July 23,
2025
3.6
Pixels, Patterns,
But No Poetry: To
See The World Like
Humans
This paper introduces LayerSkip, a method that dynamically skips
Transformer layers during inference using trainable binary masks,
significantly reducing latency without retraining. Built upon quantized
models like INT4, LayerSkip learns which layers are essential per input,
achieving up to 1.78× speedup with minimal accuracy loss across
benchmarks like MMLU, GSM8K, and HumanEval. It is compatible with
existing optimizations such as speculative decoding and early exit. Unlike
previous static or retraining-heavy methods, LayerSkip balances efficiency
and performance in real-world LLM deployment, providing a hardware-
friendly, input-adaptive solution for fast, cost-effective inference.
By Hongcheng
Gao, et al.
🔗
July 21,
2025
3.7
PyVision
Introduces Python-
Centric Framework
Where AI Writes
Tools as It Thinks
A new paper introduces PyVision, a Python-centric framework that enables
AI agents to "think by coding" in real-time. Instead of relying on pre-defined
toolkits or manual API integrations, PyVision allows LLMs to write, execute,
and reuse Python code as they reason—essentially generating tools
dynamically during problem-solving. The framework includes a code-writing
By Nikhil 🔗 July 23,
2025

policy, execution engine, and memory module that supports hierarchical
reasoning and long-term memory reuse. PyVision outperforms traditional
LLM agents on challenging tasks such as logic games, API navigation, and
coding benchmarks, highlighting the potential of code-as-reasoning for
more robust AI autonomy.
3.8
Benchmarking
GPT-4o’s Visual
Abilities Reveals
Performance Gap in
Fine-Grained Tasks
A new study benchmarks GPT-4o and other Multimodal Foundation Models
(MFMs) across 16 visual tasks, uncovering significant performance
limitations in fine-grained vision understanding. While GPT-4o excels at
OCR and captioning, it lags behind open-source models in tasks like object
localization and attribute recognition. The study reveals that GPT-4o’s
performance is inconsistent across prompt styles and lacks robustness in
real-world visual reasoning. Researchers emphasize the need for better
transparency and evaluation standards for multimodal LLMs, as current
methods often overlook their visual weaknesses.
By Sana
Hassan
🔗 July 23,
2025
3.9
Amazon Unveils
MITRA to Enhance
Tabular ML with
Synthetic Priors
Amazon researchers have introduced MITRA, a novel framework that
enhances tabular machine learning by using synthetic priors derived from
LLM-generated datasets. MITRA improves performance by generating
semantically consistent synthetic tables to pretrain models, capturing real-
world tabular distributions. When fine-tuned on target datasets, MITRA
models outperform traditional tabular models like XGBoost and TabNet
across various benchmarks. This approach bridges the data scarcity gap in
structured domains, offering a scalable solution where large annotated
tables are limited.
By Asif Razzaq 🔗 July 23,
2025

3.10
Hugging Face
Releases LoRA-
Fast: 3x Faster
Adapter Training
for LLMs
Hugging Face has introduced LoRA-Fast, a new implementation of Low-
Rank Adaptation (LoRA) that achieves up to 3× faster training speeds while
maintaining full compatibility with existing LoRA checkpoints. Built with
PyTorch 2.0, FlashAttention-2, and Triton, LoRA-Fast optimizes memory
usage and compute efficiency, significantly reducing fine-tuning time on
large models like LLaMA-3 and Mistral. Benchmarks show up to 60%
reduction in GPU memory usage and substantial FLOPs improvements.
This advancement streamlines parameter-efficient fine-tuning for
researchers and developers using modern hardware.
By Sayak Paul
and Benjamin
Bossan
🔗 July 23,
2025
3.11
Hugging Face
Launches
TimeScope: A
Benchmark for
Video LLMs with
Temporal
Reasoning
Hugging Face has released TimeScope, the first benchmark suite designed
to evaluate temporal reasoning in video-capable LLMs. The benchmark
includes 1,100+ human-annotated multiple-choice questions across four
categories—spatial, short-term, long-term, and forecasting—using
YouTube videos. TimeScope reveals significant performance gaps: even
top models like GPT-4o and Gemini 1.5 Pro struggle with temporal
understanding, scoring well below human accuracy. This initiative
spotlights the current limitations of video LLMs and provides a standardized
testbed to drive progress in multimodal temporal cognition.
By Orr Zohar et
al.
🔗 July 23,
2025
3.12
Microsoft Proposes
Ladder of
Reasoning to
Evaluate LLM
Imagination
Capabilities
Microsoft Research has introduced the Ladder of Reasoning, a novel
framework to assess how well LLMs can simulate and reason through
imagined scenarios. The benchmark includes tasks across levels of
abstraction—ranging from perception to hypothetical thinking—to test
"imaginative reasoning." Findings show that current top-tier models like
GPT-4 and Claude 3 Opus perform well on basic levels but falter at deeper,
imaginative tasks requiring abstraction and long-horizon reasoning. This
By Rachel
Lawrence
🔗 July 23,
2025

work highlights key limitations in LLMs’ generalization and proposes future
paths toward more cognitively aligned AI.
3.13
AI Coding
Challenge Exposes
Gaps in LLMs’ Real-
World Software
Engineering Skills
The inaugural AI Coding Challenge, created by a coalition of academics
and engineers, has published its first results—revealing that even leading
LLMs struggle with realistic software development tasks. Models like GPT-
4 and Claude 3 Opus underperformed when faced with multi-step
debugging, unfamiliar libraries, and open-ended coding goals. Unlike
benchmark datasets, the challenge emphasized practical coding scenarios
over synthetically curated prompts. The findings suggest that despite LLMs’
strong coding benchmarks, they fall short in authentic engineering
workflows, underscoring the need for more grounded evaluations.
By Russell
Brandom 🔗 July 23,
2025
3.14
Microsoft Proposes
Framework for
Classifying Human-
AI Interaction at
Scale
Microsoft Research has introduced a technical framework for classifying
human-AI interactions across diverse contexts and modalities. The system
distinguishes between types of human input, AI output, and task goals using
a structured labeling ontology. It supports scalable annotation of large
datasets from real-world usage, enabling systematic evaluation of AI
behavior and user intent. By standardizing how interactions are understood,
Microsoft aims to improve alignment, trust, and oversight in AI
deployments—particularly as LLMs are integrated into everyday tools and
workflows.
By Amber
Hoak,et al.
🔗 July 23,
2025
3.15
TTS-VAR: A Test-
Time Scaling
Framework for
Visual Auto-
TTS-VAR introduces the first general test-time scaling (TTS) framework for
visual autoregressive (VAR) models, casting generation as a path-search
problem. To optimize the trade-off between compute and exploration, it
uses an adaptive descending batch size during causal generation. At
coarse scales, it applies clustering-based diversity search, grouping
By Zhekai
Chen, et al. 🔗 July 24,
2025

Regressive
Generation
structural features to maintain sample variety. At fine scales, it uses
resampling-based potential selection, scoring candidates via reward
functions built from multi-scale history. Evaluated on the VAR model Infinity,
TTS-VAR boosts GenEval performance by 8.7% (from 0.69 to 0.75).
Insights show early structural cues strongly influence final quality, and
resampling effectiveness varies across scales
3.16
TeEFusion:
Blending Text
Embeddings to
Distill Classifier-
Free Guidance
TeEFusion introduces an efficient distillation method that embeds classifier-
free guidance directly into text embeddings to eliminate extra forward
passes. Unlike traditional CFG, which requires separate conditional and
unconditional inference, TeEFusion fuses those embeddings through
simple linear operations. This allows the student model to learn from the
teacher’s sophisticated sampling strategy without adding parameters or
complexity. When applied to advanced models like SD3, it achieves
inference speeds up to six times faster than the original, while preserving
comparable image quality. TeEFusion provides a streamlined path to
faster, effective text-to-image generation.
By Minghao Fu,
et al.
🔗 July 24,
2025
3.17
Anthropic Unveils
Auditing Agents to
Detect AI
Misalignment
Behaviors
Anthropic has introduced auditing agents, AI systems designed to test
other AI models for deceptive or misaligned behaviors. These agents
simulate adversarial probes—posing misleading prompts or exploring
edge-case scenarios—to identify when a model may act contrary to
intended goals. The approach helps uncover "sleeper" behaviors in frontier
models that might not surface under standard evaluations. Anthropic
frames this as a scalable strategy for AI safety testing, pushing beyond
passive evaluations to proactive red-teaming using automated AI auditors.
By Emilia David 🔗 July 24,
2025

3.18
NVIDIA Doubles
PyTorch Inference
Speed for Diffusion
Models with Torch-
TensorRT
NVIDIA has announced major performance gains for PyTorch-based
diffusion models using Torch-TensorRT, achieving up to 2x faster inference
with minimal accuracy loss. By optimizing key components like attention,
LayerNorm, and grouped convolutions, Torch-TensorRT boosts execution
on NVIDIA GPUs without requiring changes to model architecture. Tests on
Stable Diffusion 1.5 and SDXL show notable speedups across A100, H100,
and RTX 4090 cards. This advancement enables faster generation times in
real-world applications like image synthesis, enhancing responsiveness for
developers building latency-sensitive AI systems.
By Adrian
Wang, et al.
🔗 July 24,
2025
3.19
NVIDIA Boosts
Real-Time Vector
Search with cuVS
Acceleration
Library
NVIDIA has introduced cuVS, a new CUDA-based library designed to
accelerate vector search for real-time indexing and retrieval tasks in AI and
recommendation systems. Built on the RAPIDS AI framework, cuVS
supports billions of vectors with high recall and low latency, optimizing ANN
algorithms like IVF-PQ and HNSW. Benchmarks show up to 6x faster
search and 4x faster indexing over traditional CPU-based solutions. The
library integrates easily with FAISS and RAFT, offering scalable GPU-
powered search capabilities crucial for LLMs, retrieval-augmented
generation (RAG), and personalized recommendations.
By Corey Nolet,
et al.
🔗 July 24,
2025
3.20
Google Introduces
MCP for Building
Agentic AI
Experiences
Google has unveiled MCP (Modern Computing Platform), a new framework
for developing agentic AI systems capable of autonomous task execution
and dynamic decision-making. MCP abstracts infrastructure complexity,
enabling developers to orchestrate tools, memory, APIs, and LLMs through
a unified architecture. It supports persistent agent state, robust error
handling, and flexible integrations—ideal for long-running, real-world
applications like travel planning or business automation. By emphasizing
By Antony Arul
and Ruben
Gonzalez
🔗 July 24,
2025

modularity and traceability, MCP aims to streamline the shift from prompt
engineering to scalable agentic design.
3.21
Google Launches
Opal: A Lightweight
Framework for
Reproducible ML
Experimentation
Google has introduced Opal, an open-source framework designed to
streamline and standardize machine learning experimentation. Built to be
lightweight and flexible, Opal focuses on reproducibility, modularity, and
scalability, helping teams manage complex ML pipelines with minimal
overhead. It supports structured experiment tracking, configuration
management, and multi-backend training across cloud and on-prem
environments. Opal integrates seamlessly with existing tools and aims to
improve collaboration across research and production teams. This launch
highlights Google’s continued push to simplify ML workflows while
maintaining robust, repeatable experimentation practices.
By Ali Modarres 🔗 July 24,
2025
3.22
New AI Architecture
Achieves 100x
Faster Reasoning
with Minimal Data
A team of researchers has developed a novel AI architecture that delivers
100x faster reasoning than current large language models while requiring
only 1,000 training examples. Unlike LLMs, the system separates
symbolic reasoning from language understanding, enabling rapid and
interpretable logical inference. It excels at tasks like solving math word
problems and answering factual queries with minimal compute. This
architecture challenges the scaling-heavy approach of today’s LLMs,
offering a lightweight, energy-efficient alternative better suited for
constrained environments and specialized reasoning applications.
By Ben Dickson 🔗 July 25,
2025
3.23
CoSyn: The open-
source tool that’s
making GPT-4V-
CoSyn is an open-source framework developed by the University of
Pennsylvania and the Allen Institute for AI to democratize vision-language
AI. Instead of relying on real-world images, CoSyn generates synthetic data
By Michael
Nuñez 🔗 July 25,
2025

level vision AI
accessible to
everyone
by having large language models write code (e.g., LaTeX, HTML), render
visuals from it, and then generate related Q&A pairs. This produces high-
quality, diverse training datasets. Models trained with CoSyn’s 400K
dataset outperform some proprietary models like Gemini 1.5 Flash and
GPT-4V in text-rich vision tasks. CoSyn enables anyone to build powerful,
context-aware vision models without needing massive proprietary datasets
or computational budgets.
3.24
Alibaba’s Qwen
Team Proposes
Group Sequence
Policy Optimization
for LLMs
Alibaba’s Qwen team has introduced Group Sequence Policy
Optimization (GSPO), a new reinforcement learning framework designed
to enhance LLM alignment. Unlike traditional token-level optimization
methods like PPO, GSPO evaluates and optimizes entire output sequences
based on grouped rewards, leading to more stable and scalable training.
The paper shows that GSPO improves response quality and reward
alignment across various open-ended tasks, outperforming PPO in human
evaluations and benchmark metrics. The approach addresses limitations in
fine-tuning long-form generations, offering a more robust method for
aligning model behavior with human preferences.
By Qwen Team 🔗 July 25,
2025
3.25
Group Sequence
Policy Optimization
Group Sequence Policy Optimization (GSPO) is a novel RL algorithm
designed for stable and efficient training of large language models,
particularly MoE architectures like Qwen3. Unlike prior methods that use
token-level importance ratios, GSPO computes importance at the
sequence level. It then applies clipping, reward shaping, and optimization
over whole-sequence likelihoods. This approach addresses and eliminates
the instability seen in token-based algorithms like GRPO, which suffer from
high training variance and eventual collapse. GSPO consistently improves
By Qwen Team,
Alibaba Inc. 🔗 July 24,
2025

stability, efficiency, and performance in reinforcement learning settings for
LLMs
3.26
Apple Introduces
Fast-VLM for Real-
Time Vision-
Language
Understanding
Apple has introduced Fast-VLM, a new vision-language model architecture
that significantly reduces computational costs while maintaining high
accuracy. By decoupling the vision and language encoders and using an
attention-based early fusion mechanism, Fast-VLM enables 2–4× faster
inference compared to traditional VL models. It also supports batch
processing with dynamic vision token allocation, optimizing performance for
real-time applications like AR and robotics. Fast-VLM matches the
performance of compute-heavy models like Flamingo with far greater
efficiency, making it suitable for on-device use in Apple’s ecosystem.
By Apple
Machine
Learning
Research
🔗 July 23,
2025
3.27
Hugging Face
Introduces Parquet-
CDC for Efficient
Streaming Dataset
Updates
Hugging Face has released Parquet-CDC, a new format for handling
streaming dataset updates efficiently in machine learning workflows. Based
on the Apache Parquet columnar format, Parquet-CDC tracks data
changes—insertions, deletions, and updates—enabling continuous dataset
refinement without full reprocessing. This innovation supports scalable
training pipelines, especially large language models that benefit from
frequent data refreshes. Integrated with Hugging Face Datasets, it
facilitates reproducibility, version control, and modularity, while lowering
compute and storage overheads in data-centric AI development.
By Krisztian
Szucs 🔗 July 25,
2025
3.28
Geometric-Mean
Policy Optimization
(GMPO)
Geometric-Mean Policy Optimization (GMPO) enhances stability in training
language models by optimizing the geometric mean of token-level rewards
instead of the traditional arithmetic mean. This approach reduces sensitivity
to outlier importance sampling ratios common in Group Relative Policy
Optimization (GRPO). Theoretical analysis and empirical evaluation across
By Yuzhong
Zhao, et al.
🔗
July 28,
2025

mathematical reasoning and multimodal benchmarks demonstrate
improved stability and performance. A 7B-parameter version, GMPO-7B,
delivers average gains of 4.1 % on mathematical tasks (e.g. AIME24,
MATH500, OlympiadBench) and 1.4 % on multimodal reasoning tasks (e.g.
Geometry3K). Code release ensures reproducibility across benchmarks.
3.29
AGENTIC
REINFORCED
POLICY
OPTIMIZATION
Large-scale reinforcement learning with verifiable rewards (RLVR) has
proven effective in guiding LLMs through single-turn reasoning tasks.
However, realistic scenarios demand seamless multi-turn interactions
involving external tools, which expose models to high post-tool-use
uncertainty. Agentic Reinforced Policy Optimization (ARPO) addresses this
by introducing an entropy-based adaptive rollout strategy, which increases
exploration at uncertain tool-call rounds, and an advantage attribution
mechanism to assign credit across branching reasoning paths. Evaluations
on 13 benchmarks covering computational, knowledge, and search
domains show that ARPO outperforms trajectory-level RL methods while
consuming only half the tool-use budget. Code and datasets are publicly
released.
By Guanting
Dong, et al.
🔗 July 26,
2025

4.1
Intuit Brings
Agentic AI to Mid-
Market, Saving
Firms 17–20 Hours
Monthly
Intuit is rolling out agentic AI solutions for mid-market businesses,
automating complex tasks like invoice processing, payroll corrections, and
tax document prep. These multi-step autonomous workflows are saving
users 17–20 hours per month, according to early customer reports. The
system leverages Intuit’s proprietary LLMs alongside accounting and
compliance logic to ensure accuracy. Designed for firms with 10–100
employees, it marks a major shift in how automation reaches small-to-
midsize enterprises—offering efficiency gains once reserved for enterprise-
scale companies.
By Sean
Michael Kerner 🔗 July 22,
2025
4.2
Delve raises $32 M
to scale AI agents
automating
regulatory
compliance across
frameworks.
On July 22, 2025, TechCrunch announced Delve, founded by 21-year-old
MIT dropouts Karun Kaushik and Selin Kocalar, closed a $32 million
Series A at a $300 million valuation led by Insight Partners. Their platform
employs AI agents to automate complex regulatory compliance—originally
focused on HIPAA, now supporting SOC 2, PCI, GDPR, ISO, and more—
by gathering evidence, generating reports, updating logs, and integrating
across internal systems. Since their seed round, Delve expanded from 100
to over 500 customers, including AI startups and enterprises. The company
aims to eliminate administrative bottlenecks and scale into adjacent
domains like cybersecurity and governance.
By Tage
Kene-Okafor 🔗 July 22,
2025
4.3
iOS 26 Beta 4
brings back AI-
driven news
notification
summaries and
refines the
Liquid Glass UI.
Apple’s iOS 26 Beta 4, released on July 22, 2025, restores its AI-powered
notification summaries for news and entertainment. After earlier backlash
over accuracy—such as misinterpreting BBC articles—summaries are now
opt-in and include disclaimers noting potential meaning changes. This
move aims to balance innovation with transparency. Alongside the AI
updates, Apple continues refining the Liquid Glass design language with
enhanced translucency and dynamic UI effects across system apps like
By Sarah Perez 🔗 July 22,
2025

Photos and Music. Improvements include navigation bar styling, splash
screen transitions, and revamped onboarding. A broader public beta rollout
is expected later in the week.
4.4
Amazon acquires
Bee to bring
wearable,
always-on voice AI
to mainstream.
On July 22, 2025, Amazon confirmed its acquisition of Bee, a San
Francisco–based startup behind an AI-enabled wristband and Apple Watch
app that records ambient audio to generate summaries, reminders, and
to-do lists. Priced at $49.99 plus a $19 monthly subscription, the device
integrates with email, calendar, contacts, and location to offer personalized
insights. Bee’s privacy stance—no audio storage and user-controlled
muting—is set to continue, with additional controls planned. Bee raised $7
million in seed funding. The acquisition signals Amazon's renewed push
into personal wearable AI, following its 2023 halo tracker exit.
By Amanda
Silberling 🔗 July 22,
2025
4.5
Cequence
Launches AI
Gateway to Secure
Agent Access to
Enterprise Apps
Cequence Security has unveiled an AI Gateway designed to manage and
secure real-time connectivity between AI agents and enterprise
applications. Announced July 22, 2025, the gateway provides
authentication, authorization, and monitoring controls to ensure safe LLM-
to-app interactions—especially for agents performing sensitive tasks like
transactions or database queries. It acts as a policy enforcement layer,
detecting anomalies and preventing misuse while maintaining performance.
As enterprises increasingly deploy AI agents, Cequence’s solution
addresses the urgent need for robust access control and governance in
multi-agent, API-rich environments.
By Duncan
Riley 🔗 July 22,
2025
4.6
Uber's PerfInsights
leverages GenAI to
automate Go
Uber's PerfInsights, developed during Hackdayz 2024, automates
performance optimization in Go services. By analyzing CPU and memory
profiles from production services, it identifies the top 30 most resource-
By Lavanya
Verma et al.
🔗 July 22,
2025

service
optimization,
reducing manual
profiling efforts.
intensive functions. The system then applies a curated catalog of
performance antipatterns, such as unbounded memory allocations and
inefficient string operations, to detect potential inefficiencies. Utilizing large
language models (LLMs), PerfInsights validates optimization suggestions,
ensuring accuracy and reducing developer uncertainty. Since its
deployment, PerfInsights has merged hundreds of diffs into Uber's Go
monorepo, transforming optimization into a scalable, repeatable practice.
4.7
SecurityPal Uses AI
and Nepal-Based
Experts to
Accelerate
Enterprise Security
Reviews
SecurityPal has developed a hybrid solution combining AI models with a
team of Nepal-based security analysts to streamline enterprise security
questionnaires—cutting turnaround times by 87x or more. Announced July
23, 2025, the system automates form parsing, pre-fills answers from
security documentation, and routes edge cases to trained experts for
review. This model ensures speed without sacrificing accuracy or
compliance, addressing a major pain point in B2B sales and vendor
onboarding. SecurityPal’s approach demonstrates how AI-human
collaboration can optimize complex enterprise workflows in regulated
industries.
By Carl Franzen 🔗 July 23,
2025
4.8
Google DeepMind
Applies AI to
Decode Ancient
Roman History
Google DeepMind is using AI to help historians better understand ancient
Roman history, applying large language models to analyze, translate, and
contextualize Latin texts and inscriptions. Announced July 23, 2025, the
initiative supports archaeologists and classicists by revealing patterns in
fragmented records, enhancing translations, and uncovering socio-political
insights. The system has been trained on curated historical corpora and is
designed to assist, not replace, human experts. This collaboration between
AI and academia showcases how language models can contribute
meaningfully to historical research and cultural preservation.
By James
Farrell 🔗 July 23,
2025

4.9
AI-Powered Ads
Fuel Global
Entertainment and
Media Growth, Says
PwC
AI-driven advertising is now the primary growth engine for the global
entertainment and media industry, according to PwC’s July 2025 report.
Personalized ad targeting, generative content creation, and real-time
optimization are enabling companies to scale campaigns more efficiently
and reach highly segmented audiences. The report projects global industry
revenue will hit $2.8 trillion by 2028, with AI technologies playing a pivotal
role in digital advertising, streaming, and immersive media. This trend
reflects a broader shift where AI is not just enhancing content—but driving
the economics behind it.
By Summer
Zhen 🔗 July 24,
2025
4.10
Alphabet Q2 2025
Earnings Show AI
Driving Revenue
Growth Across
Google Products
Alphabet’s Q2 2025 earnings reveal that AI integration is boosting
revenue across Google’s core products, including Search, Ads, and Cloud.
CEO Sundar Pichai highlighted advances in generative AI as key to
enhancing user experience and advertiser performance. Gemini-powered
features now appear in over 60% of Search queries and are improving click-
through rates. Google Cloud revenue also surged due to increased demand
for AI infrastructure and model deployment. The results underscore how
deeply AI is embedded in Alphabet’s strategy, reinforcing its position in the
competitive enterprise AI landscape.
By The Verge 🔗 July 24,
2025
4.11
SyncoGen Unveils
ML Framework for
Synthesizable 3D
Molecular
Generation
Researchers have introduced SyncoGen, a machine learning framework
designed to generate synthesizable 3D molecular structures by jointly
modeling molecular graphs and atomic coordinates. Unlike traditional
methods that handle structure and chemistry separately, SyncoGen fuses
graph neural networks and geometric modeling to co-optimize both
synthesis viability and spatial realism. It significantly outperforms prior
models on key benchmarks for molecular validity, novelty, and drug-
By Sajjad Ansari 🔗 July 23,
2025

likeness. This breakthrough enables more accurate molecule discovery in
pharmaceutical R&D, offering a scalable path for AI-driven drug design.
4.12
DraftWise and
Cohere Partner to
Transform Legal
Drafting with AI
Legal tech startup DraftWise has partnered with Cohere to integrate its
RAG-based Command R+ model into contract drafting workflows,
significantly improving legal document creation and review. The AI retrieves
relevant clauses, legal precedents, and language patterns from a firm’s
proprietary data, offering real-time, context-aware suggestions. This
minimizes errors and boosts lawyer productivity without replacing expert
judgment. Already adopted by top-tier firms like Latham & Watkins and
Orrick, DraftWise demonstrates how AI can securely accelerate high-
stakes legal processes while maintaining firm-level data privacy.
By Cohere
Team 🔗 July 23,
2025
4.13
NVIDIA Unveils AI
Tools for
Personalized Ads
and 3D Content
Generation
NVIDIA has introduced a suite of AI tools designed to personalize
advertising and accelerate 3D content generation for digital marketing.
Leveraging NVIDIA’s Edify and Picasso models, the tools enable brands to
create customized product images, videos, and interactive experiences at
scale. The platform uses generative AI to adapt content to user
preferences, environments, and behaviors, enhancing engagement while
reducing creative costs. It also supports real-time A/B testing and iterative
refinement, marking a major step toward hyper-personalized, immersive ad
experiences.
By James Mills 🔗 July 23,
2025
4.14
NVIDIA Proposes
Advanced Pipeline
for Accurate PDF
Data Extraction
NVIDIA has detailed a multi-stage pipeline for extracting structured
information from complex PDF documents, addressing challenges in layout
variance, font noise, and embedded graphics. The proposed system
combines OCR, layout parsing, LLM-based table detection, and knowledge
graph construction to transform PDFs into machine-readable data for
By Raja Biswas
and Bo Liu
🔗 July 23,
2025

downstream applications like search and analytics. It outperforms
traditional methods by preserving context, structure, and semantic
meaning—especially in scientific, financial, and legal documents. This
approach enhances enterprise workflows dependent on unstructured or
semi-structured documents.
4.15
Google Photos
Adds AI Features
for Stylized
Remixing and Video
Generation
Google Photos has introduced new AI-powered tools that let users remix
their images into various artistic styles and generate short videos from static
pictures. The update includes style transfer models, animated transitions,
and contextual scene generation to create personalized visual stories.
Users can apply cinematic effects or reimagine photos as watercolors,
sketches, and more. The features are designed to enhance user creativity
and engagement while showcasing Google’s advances in generative visual
AI. Rollout is expected globally over the coming weeks.
2025
4.16
YouTube Shorts
Adds Image-to-
Video AI Tool and
Generative Effects
YouTube Shorts is rolling out new generative AI features, including an
image-to-video tool that transforms still photos into short animated clips
using motion prediction and scene generation models. Users can also apply
new AI effects like stylized backgrounds and dynamic overlays to enhance
video creativity. The update aims to empower creators with fast, accessible
content tools while competing with TikTok and Instagram Reels. These
features are part of Google’s broader strategy to embed generative AI
across its media ecosystem.
By Aisha Malik 🔗 July 23,
2025
4.17
Proton Launches
Privacy-First AI
Assistant with End-
to-End Encryption
Proton, known for its secure email and VPN services, has launched a
privacy-focused AI assistant that encrypts all conversations end-to-end and
stores no chat logs. The assistant runs on Proton’s in-house infrastructure
with zero third-party model access, aligning with its commitment to user
By Ivan Mehta 🔗 July 22,
2025

privacy. Unlike mainstream assistants, Proton’s tool is designed for
sensitive tasks like legal queries or confidential planning. It supports
multiple languages and offers concise, secure summaries and
suggestions—pioneering a new standard for privacy in consumer AI
applications.
4.18
Captain Cinema:
Towards Short
Movie Generation
Captain Cinema is a framework for generating coherent short movies from
textual story descriptions. It first uses top-down keyframe planning to create
a narrative-aligned sequence of key scenes, ensuring long-range visual
and storyline consistency. These keyframes feed into a bottom-up video
synthesis model, conditioned to learn long-range context and generate
seamless dynamics between frames. The system is trained with an
interleaved training strategy applied to Multimodal Diffusion Transformers
(MM-DiT), optimized for long-context cinematic data. Experiments on a
specialized cinematic dataset show Captain Cinema efficiently produces
narrative-consistent, visually high-quality short films
By Junfei Xiao,
et al. 🔗 July 24,
2025
4.19
Freed’s AI Medical
Scribe Reaches
20,000 Clinicians
Amid Growing
Competition
Freed, a medical AI transcription startup, reports that over 20,000 clinicians
now use its voice-based scribe tool to automate clinical note-taking. The AI
transcribes patient visits in real time and structures the data into EHR-
compatible formats, reducing administrative burden. Freed claims improved
accuracy and compliance through in-house models and data pipelines.
However, competition is intensifying as rivals like Nabla and Abridge
expand rapidly with similar offerings, signaling a booming but crowded
market for AI-assisted healthcare documentation.
2025
4.20
Chime backer
Lauren Kolodny
Lauren Kolodny, founding partner at Acrew Capital and known as an early
investor in Chime, has now led a $20 million Series A investment in Alix, a
By Marina
Temkin 🔗 July 24,
2025

bets on AI to
revolutionize estate
processing
startup aiming to transform inheritance and estate processing through
artificial intelligence. Alix accelerates estate management by automating
processes such as data extraction from documents, form completion, and
communication with financial institutions. The idea for the company was
born from founder Alexandra Mysoor's own exhausting 18-month
inheritance experience. Alix provides its services at a cost ranging from
$9,000 to $12,000 for a typical user. According to Kolodny, this model aims
to make technology accessible to a broader audience by simplifying
complex financial transactions.
4.21
Leena AI Launches
Voice-Enabled AI
Colleagues for
Enterprise
Workflows
Leena AI has unveiled a new generation of voice-enabled AI colleagues
designed to collaborate with human employees across HR, IT, and
operations. These agents can attend meetings, take voice commands,l
manage workflows, and respond to natural-language queries in real time.
Built on proprietary LLMs and integrated with enterprise tools, they aim to
boost productivity by handling routine tasks and information retrieval.
Leena’s offering reflects a broader trend of embedding conversational AI
agents directly into enterprise environments to support day-to-day
operations.
By KYT 🔗 July 24,
2025
4.22
Industrial AI Startup
Copia Gains
Traction by
Promising Long-
Term Independence
Industrial AI startup Copia Automation is attracting manufacturers by
pledging it won’t be acquired, positioning itself as a stable, long-term
partner in a volatile market. Copia offers GitHub-like version control tools
for industrial automation software, addressing critical pain points in
manufacturing systems. By emphasizing independence, Copia reassures
clients wary of disruptions from corporate acquisitions. With funding from
firms like Lux Capital, Copia is scaling its customer base among industrial
giants seeking reliability and innovation. The startup’s stance reflects a
By Sean O'Kane 🔗 July 24,
2025

growing desire for trustworthy AI partners in complex, high-stakes
environments.
4.23
Google Tests “Web
Guide” AI Search
Experience for
Organized Topic
Exploration
Google is piloting a new AI-driven search feature called Web Guide,
designed to help users explore broad topics through structured,
summarized content. When users search general topics like “climate
change,” Web Guide presents subtopics, key questions, and AI-generated
summaries based on real-time web content. The feature aims to reduce
overwhelming information by organizing results into digestible sections,
improving discovery and learning. Still in experimental rollout on mobile in
the U.S., Web Guide reflects Google’s push to modernize search by
integrating generative AI for more intuitive and educational experiences.
2025
4.24
Samsung Invests in
Irreverent Labs’ AI
to Analyze Massive
Video Libraries
Samsung Next has invested in Irreverent Labs, a startup developing AI
models that can analyze and understand vast volumes of video content—
thousands of hours at once. Unlike traditional models trained on short clips,
Irreverent’s technology enables higher-level video understanding, such as
summarization, search, and content indexing. The startup aims to support
industries like security, sports, and entertainment where scalable video
analysis is critical. The investment reflects growing demand for AI tools that
process unstructured video data efficiently and highlights Samsung’s
interest in next-gen multimodal AI capabilities.
2025
4.25
LegalOn Raises
$50M to Expand AI
Tools for In-House
Legal Teams
Legal tech startup LegalOn has secured $50 million in a Series E round led
by SoftBank to scale its AI-driven legal document analysis platform.
Focused on serving in-house legal teams, LegalOn’s tools help automate
contract review, compliance checks, and legal risk detection using
proprietary AI trained on U.S. and Japanese law. The funding will support
By Kate Park 🔗 July 24,
2025

U.S. expansion and deepen product capabilities tailored to corporate legal
departments. As legal workflows grow more complex, LegalOn’s rise
reflects increasing demand for domain-specific AI that can improve
efficiency and accuracy in enterprise legal operations.
4.26
Google Launches
AI-Powered Virtual
Try-On for Clothing
Searches
Google has rolled out an AI-powered virtual try-on feature for clothing in
mobile search, allowing users to see how garments look on a diverse range
of real human models. Available for over 60 brands, the feature uses a
diffusion-based generative model to render clothes realistically on different
body types, poses, and skin tones. Shoppers can adjust results by size and
fit preferences. Initially live in the U.S., it marks a major step in AI-enhanced
retail search, merging fashion with personalization to reduce uncertainty in
online shopping.
By Aisha Malik 🔗 July 24,
2025
4.27
MIT Launches
ChemXploreML to
Predict Chemical
Properties with AI
MIT researchers have introduced ChemXploreML, a web app that
leverages machine learning to predict key chemical properties from
molecular structures. Designed for scientists and educators, the tool
enables users to input molecules via drawing or SMILES notation and
instantly access predictions for solubility, boiling points, and other physical
properties. The app uses models trained on extensive public databases and
provides confidence scores for results. By accelerating early-stage
research and reducing reliance on physical testing, ChemXploreML
showcases how AI can streamline material discovery and education in
chemistry.
By Danielle
Randall
Doughty
🔗 July 24,
2025

4.28
Cohere Proposes
Secure and
Compliant AI
Strategy for Europe
Cohere has released a detailed framework outlining its approach to secure
and responsible AI deployment in Europe, aligning with the EU AI Act and
GDPR. The strategy emphasizes model transparency, robust evaluations,
and strong data governance—including EU-specific model hosting and fine-
tuning controls. Cohere commits to on-premise and private cloud
deployment options to support compliance and sovereignty. The plan also
outlines practices for mitigating misuse and bias in AI systems. This
initiative reflects growing industry momentum toward region-specific AI
compliance and ethical alignment with European regulatory expectations.
By Cohere
GAAP 🔗 July 24,
2025
4.29
Anthropic Details
Internal Use of
Claude for Secure
Code Assistance
Anthropic has shared insights into how its teams use Claude as a secure
and collaborative coding assistant. Developers rely on Claude for code
explanation, debugging, refactoring, and documentation—particularly
valuing its ability to handle large context windows for navigating complex
systems. The company ensures internal data privacy by using isolated
environments with no external API calls. Claude also helps cross-functional
teams—like product and legal—engage with codebases, enhancing
technical alignment. The post highlights how AI can boost productivity and
cross-team understanding in sensitive, high-stakes development
workflows.
By Anthropic
Newsroom 🔗 July 24,
2025
4.30
Stanford HAI
Launches In Silico
Center to Advance
Computational
Social Science
Stanford HAI has launched the In Silico Center, a pioneering research
initiative leveraging generative agents and large language models to
simulate complex social behaviors at scale. The center aims to
revolutionize social science by enabling experiments traditionally limited by
cost or feasibility to be conducted digitally. Initial studies include modeling
misinformation spread, political polarization, and economic mobility. By
creating lifelike, interactive AI agents, researchers can test interventions
By Katharine
Miller 🔗 July 25,
2025

and policies in controlled virtual environments, offering a powerful new
method for understanding human societies.
4.31
Hugging Face
Unifies Developer
Tools with New
huggingface CLI
Hugging Face has launched a new unified command-line interface
(huggingface CLI) to streamline interaction with its platform. Consolidating
multiple tools into a single CLI allows developers to manage models,
datasets, spaces, and APIs more efficiently. Key features include
environment management, fine-tuning orchestration, Git-like versioning,
and dataset exploration—all within one terminal tool. Designed for
productivity and scalability, the CLI supports both beginner and enterprise
users developing AI workflows, reducing friction across experimentation,
deployment, and collaboration.
By Lucain
Pouget et al. 🔗 July 25,
2025
4.32
Google Tests Opal:
An AI-Powered
“Vibe” Coding App
for Creative
Development
Google is testing Opal, an experimental coding app aimed at blending
creativity and AI-assisted development. Unlike traditional IDEs, Opal
focuses on the “vibe” or intent behind code, offering an ambient,
collaborative interface for building web apps and prototypes. It uses AI to
understand loosely defined user goals and generate or modify code
accordingly. Currently available to a small group of testers, Opal signals
Google's exploration of more intuitive, expressive programming
environments that lower the barrier to entry for non-traditional developers.
2025
4.33
AI Referrals to Top
Websites Surge
357% YoY, Topping
1.13B in June
According to new data from Similarweb, AI-driven web referrals reached
1.13 billion in June 2025—a 357% year-over-year increase. ChatGPT led
the trend, generating 2.6 times more traffic than Google Gemini, followed
by Perplexity and Copilot. The spike highlights the growing influence of AI
assistants as discovery engines, shifting user behavior from search to
conversational interfaces. Most referrals targeted educational, tech, and
2025

news domains. The report underscores AI’s emerging role as a top-of-
funnel driver, signaling a reshaping of how users find and access
information online.
4.34
The Web Browser
Evolves into AI-
Powered Agent,
Redefining Search
and Navigation
VentureBeat reports on a major shift in how browsers function, evolving into
AI agents that perform tasks rather than just display links. New AI-native
browsers like Arc, Perplexity, and Google's Search Generative Experience
now summarize content, execute commands, and deliver direct answers.
These agents reduce the need for scrolling and link-clicking, acting more
like proactive assistants. This evolution transforms the browser into an
intelligent interface, reshaping search behavior and creating new design
paradigms for the web’s AI-driven future.
By Taryn Plumb 🔗 July 28,
2025
4.35
E2B Raises $21M as
Its AI-Powered Dev
Environments Gain
Traction in Fortune
100
AI startup E2B has secured $21 million in Series A funding as its cloud-
based development environments now serve 88% of Fortune 100
companies. E2B offers programmable, ephemeral environments optimized
for AI workloads and continuous deployment, simplifying infrastructure for
AI app development. Its environments integrate directly with APIs and agent
frameworks, accelerating developer productivity and collaboration. The
funding, led by Khosla Ventures, will support product scaling and growth
across enterprise sectors. E2B’s rise reflects growing demand for AI-native
developer tooling.
By Michael
Nuñez
🔗 July 28,
2025
4.36
Microsoft
Transforms Edge
into AI Agent with
New Copilot Mode
Microsoft is turning its Edge browser into an AI-powered assistant with the
launch of a new Copilot mode. The feature enables Edge to summarize
webpages, automate form-filling, compare product data, and even execute
tasks like travel planning or online purchases. Integrated deeply into
Windows, Copilot leverages Microsoft’s Prometheus model and Bing Chat
By Duncan
Riley
🔗 July 28,
2025

to act more like a digital agent than a traditional browser tool. This shift
aligns with the industry trend of making browsers active participants in
workflows, not just portals.
4.37
Citi Expands
Research Coverage
into Private Tech
Firms Using AI
Tools
Citigroup has announced a major expansion of its research division to
include private, mostly tech-focused companies—many of which are early-
stage and pre-IPO. The initiative leverages AI-powered tools to analyze
opaque data sources such as hiring trends, website traffic, and alternative
signals to produce investor-grade insights. This move addresses growing
demand for intelligence on high-growth private markets, particularly in
sectors like AI and fintech. It also reflects Wall Street’s increasing reliance
on AI to unlock value in undercovered spaces.
By Reuters 🔗 July 29,
2025
4.38
Auterion to Supply
Ukraine with 33,000
AI-Powered Drone
Guidance Kits
Swiss-American drone software company Auterion will deliver 33,000 AI-
guided drone navigation kits to Ukraine, enhancing its autonomous strike
and surveillance capabilities. The kits feature edge AI processing and are
compatible with various commercial drones, enabling real-time target
recognition and autonomous pathfinding in GPS-denied environments. The
deployment reflects a shift toward low-cost, scalable AI-powered warfare
tools, allowing Ukraine to repurpose consumer drones for military use at
scale. This development highlights the accelerating role of AI in modern
conflict zones.
2025

5.1
OpenAI and Oracle
Partner to Build 4.5
GW of New AI Data
Center Capacity
OpenAI has teamed up with Oracle to develop 4.5 gigawatts of new AI
data center capacity in the United States, one of the largest infrastructure
expansions in generative AI to date. The partnership will support growing
compute demands driven by advanced models and autonomous AI
agents. Oracle will provide cloud infrastructure through Oracle Cloud
Infrastructure (OCI), with an emphasis on renewable energy sources. The
expansion highlights the escalating race for compute power among tech
giants and reinforces Oracle’s position as a key AI infrastructure provider
for OpenAI’s rapidly scaling operations.
By Maria
Deutscher
🔗 July 22,
2025
5.2
AWS Expands
Generative AI
Innovation Center
with $100M
Investment Boost
AWS is expanding its Generative AI Innovation Center with a new $100
million investment, aiming to accelerate enterprise adoption of
generative AI across healthcare, financial services, and manufacturing.
Announced on July 22, 2025, the expansion will fund additional AI experts,
solution architects, and domain specialists to help clients build and deploy
tailored generative AI applications. Since its launch in 2023, the center has
supported over 1,000 organizations. This investment reflects AWS’s
continued push to scale AI consulting and infrastructure as demand for
custom, domain-specific generative AI solutions grows rapidly.
By Zeus
Kerravala
🔗 July 22,
2025
5.3
Amazon Shuts Down
Shanghai AI
Research Lab Amid
Shifting Global
Strategy
Amazon has shuttered its Shanghai AI research lab, reassigning staff
and projects as part of a strategic pivot away from China-based AI
development. Reported by the Financial Times on July 23, 2025, the
closure affects teams working on Alexa and foundational AI models,
and reflects growing U.S. corporate caution amid rising geopolitical and
regulatory tensions. Amazon will concentrate future AI investments in the
U.S., Canada, and Europe. The move mirrors broader industry trends as
By Financial
Times
🔗 July 23,
2025

tech giants restructure global R&D footprints to align with national security
and policy concerns.
5.4
White House to
Promote U.S. AI
Globally, Push Back
on Restrictive
Foreign Regulations
The White House is set to unveil a strategy aimed at promoting U.S. AI
technologies abroad while opposing foreign regulations it deems overly
restrictive. According to a document reviewed by Reuters (July 22, 2025),
the plan supports global AI export initiatives and encourages allies to
adopt “innovation-friendly” governance, countering China’s and the EU’s
stricter regulatory models. The effort includes support for U.S. firms facing
trade barriers and a push for harmonized global standards. It reflects a
broader geopolitical strategy to maintain U.S. leadership in AI through
diplomacy and economic policy.
2025
5.5
Mistral AI pioneers
transparent
environmental
impact reporting
with a full lifecycle
analysis of its
models.
Mistral AI released a comprehensive lifecycle analysis (LCA) of its large
language models, including Mistral Large 2, assessing greenhouse gas
emissions, water use, and resource depletion. The study found training
Mistral Large 2 produced 20.4 kilotons of CO₂ equivalent and consumed
281,000 cubic meters of water. Inference for a typical 400-token AI
assistant reply emits 1.14 grams of CO₂ equivalent and uses 45 milliliters
of water. This initiative aims to set a global environmental standard for AI,
promoting transparency and sustainability. Mistral advocates choosing
appropriately sized models to reduce environmental impact, aligning with
broader sustainable AI efforts.
By Mistral AI 🔗 July 22,
2025
5.6
Taiwan Unveils AI
Initiative to Boost
Economy by $510
Billion
Taiwan has launched a sweeping AI development strategy aimed at
generating $510 billion in economic output over the next five years.
Announced on July 23, 2025, the plan includes investments in smart
manufacturing, semiconductor innovation, and AI workforce training.
2025

Premier Cho Jung-tai emphasized Taiwan’s goal to become a global AI
hub, leveraging its strength in chips and electronics. The initiative includes
regulatory updates and public-private collaboration to drive AI adoption
across industries. This marks one of Asia’s most ambitious national AI
programs to date.
5.7
Former Anthropic
Exec Raises $15M to
Launch AI Agent
Insurance and Safety
Platform
A former Anthropic executive has raised $15 million in seed funding to
launch a new startup focused on insuring and securing AI agents.
Announced July 23, 2025, the platform offers liability insurance, safety
tooling, and deployment guidelines to help startups mitigate risks from
autonomous AI agents. It targets emerging use cases where AI systems
operate independently across sensitive workflows. The initiative reflects
growing demand for AI-specific risk management infrastructure,
echoing developments in cybersecurity and enterprise compliance.
Backers include top VC firms and AI safety advocates.
By Michael
Nuñez
🔗 July 23,
2025
5.8
White House plan
signals “open-weight
first” era—and
enterprises need
new guardrails
The White House’s new AI Action Plan signals strong federal support for
open-weight AI models, encouraging transparency and innovation. While
the plan applies directly to government agencies, it’s expected to influence
enterprise AI strategy broadly. It prioritizes infrastructure investment,
open-source collaboration, and evaluation of foreign AI risks. By promoting
access to model weights, the plan shifts the balance away from closed
systems like GPT-4. However, it also calls for new safeguards to manage
risks in enterprise use. Experts say clearer guidance is needed as open-
weight models reshape industry norms and governance expectations.
2025

5.9
Holistic AI Report:
Red Teaming Could
Have Averted Grok-
4’s Public Meltdown
A new report from Holistic AI argues that robust red teaming and system
testing could have prevented the public failure of xAI’s Grok-4 model,
which recently drew backlash for generating antisemitic and extremist
content. Published July 23, 2025, the analysis faults insufficient pre-
deployment safeguards and highlights the need for iterative stress testing,
bias audits, and safety fine-tuning. The report calls for industry-wide
standards for safety validation and deployment readiness, especially
for AI agents operating in public-facing or high-impact contexts. Grok-4’s
case is now seen as a cautionary tale.
By KYT 🔗 July 23,
2025
5.10
Amazon’s NOVA AI
Challenge Focuses
on Real-World
Attacks and Secure
AI Development
Amazon has launched the NOVA AI Security Challenge, inviting
researchers to stress-test AI systems against real-world threats like
prompt injection, model evasion, and data leakage. Announced July 23,
2025, the competition aims to advance secure-by-design practices for
AI development, mirroring bug bounty programs in cybersecurity.
Participants will test models in sandboxed environments, with top findings
informing AWS’s AI safety protocols. The initiative reflects growing
industry consensus on the need for proactive red-teaming and
transparency to prevent misuse as AI agents and assistants gain
autonomy and access to sensitive systems.
By Duncan
Riley
🔗 July 23,
2025
5.11
Anthropic Endorses
U.S. AI Action Plan,
Urges Focus on
Safety, Red-
Anthropic has responded positively to the Biden Administration’s U.S. AI
Action Plan, praising its focus on safety, security, and innovation. The
company emphasized the importance of independent red-teaming,
standardized evaluations, and advanced AI safety research. It also
By Anthropic
Newsroom 🔗 July 23,
2025

Teaming, and Global
Standards
advocated for international coordination on AI governance, aligning with
the U.S. proposal for global alignment through the G7 and UN. Anthropic
highlighted the need for rigorous frameworks that scale with model
capabilities to ensure responsible development as AI progresses toward
frontier systems.
5.12
Anthropic and
UChicago Launch
Research Program
on AI’s Economic
Impact
Anthropic has partnered with the University of Chicago’s Becker Friedman
Institute to launch a multi-year research initiative examining AI’s long-term
effects on labor, productivity, and economic growth. The program will
combine theoretical and empirical approaches to assess how frontier AI
systems reshape industries and labor markets. It aims to produce
actionable insights for policymakers, academics, and businesses
navigating economic transitions driven by rapid AI deployment. This
collaboration underscores the growing emphasis on evidence-based
strategies to manage AI's transformative economic potential.
By Anthropic
Newsroom
🔗 July 23,
2025
5.13
Trump’s “Anti-Woke
AI” Order Aims to
Regulate Training
Data in U.S. Tech
Former President Donald Trump has issued an executive order targeting
“woke bias” in AI, mandating U.S. tech companies to disclose and adjust
the data used to train AI systems. The order calls for transparency in model
inputs, aims to ban perceived political bias, and directs federal agencies
to restrict procurement of AI tools that don’t comply. Critics warn it could
politicize AI development and conflict with open-source practices.
Supporters argue it protects free speech and viewpoint diversity in
algorithmic outputs.
Source: TechCrunch (23 July 2025), authored by Devin Coldewey.
By Rebecca
Bellan
🔗 July 23,
2025

5.14
Trump’s AI Action
Plan Proposes
Blocking Chip
Exports to China
Donald Trump’s AI Action Plan includes a proposed ban on advanced AI
chip exports to China, citing national security and economic
competitiveness. The plan aligns with previous export controls but
expands their scope to cover a wider range of semiconductors and
manufacturing tools. However, critics note the policy lacks technical
specifics, enforcement mechanisms, and coordination with allies. Industry
leaders express concern about potential retaliation and supply chain
disruptions. The proposal signals a continued hardline stance on China’s
AI development under a possible second Trump administration.
By Rebecca
Szkutak
🔗 July 23,
2025
5.15
Trump’s AI Strategy
Prioritizes
Deregulation to
Accelerate U.S.
Innovation
Donald Trump’s new AI strategy emphasizes reduced regulation to
stimulate rapid innovation and maintain U.S. dominance over China. The
plan proposes slashing compliance burdens, limiting federal oversight,
and rolling back existing AI safety mandates in favor of “freedom to
innovate.” It also includes tax incentives for AI R&D and support for
domestic chip production. Critics argue the approach weakens ethical
safeguards and overlooks long-term risks, while supporters see it as a pro-
business push to counter China’s state-backed AI growth.
By Rebecca and
Bellan Maxwell
Zeff
🔗 July 23,
2025
5.16
Trump to Unveil AI
Roadmap Focused
on Deregulation,
National Security,
and China
Donald Trump is set to unveil a comprehensive AI Roadmap emphasizing
deregulation, AI chip export bans to China, and reduced government
oversight. The plan supports American AI leadership through tax breaks,
innovation zones, and streamlined compliance for startups. It also
introduces mandates to prevent “woke bias” in AI and expands restrictions
on advanced semiconductor exports. While the strategy positions the U.S.
in direct competition with China’s AI ambitions, critics warn of weakened
ethical oversight and unclear implementation pathways.
By Maxwell Zeff 🔗 July 23,
2025

5.17
Trump Considered
Breaking Up NVIDIA
in AI Action Plan,
Citing Market Power
During his AI Action Plan speech, Donald Trump revealed that his team
had considered breaking up NVIDIA, citing its dominant position in AI chip
manufacturing. While no formal antitrust action was announced, the
mention signals growing political scrutiny over NVIDIA’s control of AI
hardware. The statement adds to Trump’s broader policy themes:
promoting domestic competition, restricting chip exports to China, and
curbing perceived monopolistic behavior in the tech sector. Analysts warn
this rhetoric could increase regulatory pressure on AI hardware giants.
By Maria
Deutscher 🔗 July 24,
2025
5.18
AI-Generated Slop Is
Undermining Bug
Bounty Programs
Security researchers and platforms are raising concerns over a surge in
low-quality, AI-generated bug reports flooding bug bounty programs.
These “AI slop” submissions often contain fabricated or irrelevant
vulnerabilities, overwhelming reviewers and reducing the effectiveness of
legitimate disclosures. Some platforms like HackerOne have already
started suspending accounts abusing AI tools. The trend highlights a
growing challenge in balancing AI assistance with accountability in
cybersecurity. Industry leaders are now calling for clearer guidelines and
vetting practices to prevent generative AI from degrading trust and
productivity in the vulnerability disclosure ecosystem.
By Lorenzo
Franceschi-
Bicchierai
🔗 July 24,
2025
5.19
Trump Campaign
Unveils Aggressive
AI Strategy Centered
on Deregulation
The Trump 2024 campaign has outlined an ambitious AI action plan
prioritizing deregulation, innovation, and national AI leadership. The
strategy opposes government overreach and calls for eliminating Biden-
era executive orders on AI governance. It promotes expanding private
sector freedom, limiting liability for AI developers, and increasing federal
investment in AI R&D. The plan also emphasizes AI in defense and border
security. Critics argue it downplays safety and ethics in favor of rapid
By Caroline
Meinhardt et al.
🔗 July 24,
2025

commercialization. The proposal reflects a broader political divide on how
the U.S. should balance AI innovation with oversight.
5.20
Meta Hires GPT-4
Co-Creator Shengjia
Zhao to Lead
Superintelligence
Labs
Meta has appointed Shengjia Zhao, former OpenAI researcher and GPT-
4 co-creator, as Chief Scientist of its Superintelligence Labs—a division
focused on building frontier AI systems. Zhao brings expertise in model
alignment, scaling laws, and advanced LLM architecture. His leadership
signals Meta’s intensified pursuit of AGI and safe superintelligence, aiming
to compete directly with OpenAI and Google DeepMind. The move
underscores growing talent wars among AI giants, as companies race to
define safe development frameworks while advancing cutting-edge
capabilities in highly autonomous, reasoning-rich systems.
2025
5.21
Anthropic Introduces
Auditing Agents to
Detect AI
Misalignment Risks
Anthropic has launched auditing agents, specialized AI systems
designed to evaluate other AI models for signs of misalignment,
deception, or harmful capabilities. These agents simulate adversarial
testing and are capable of detecting concealed behaviors that might only
emerge under specific conditions. The initiative reflects Anthropic’s
broader strategy to build “constitutional AI” through robust evaluation and
oversight. By integrating auditing agents into model development,
Anthropic aims to identify risks before deployment, setting a new standard
for internal safety auditing of frontier models.
By Emilla David 🔗 July 24,
2025
5.22
Trump Reportedly
Pauses AI Chip
Export Controls to
Boost China Trade
Deal
According to the Financial Times, former President Donald Trump has
privately signaled plans to pause AI chip export controls to China if re-
elected, aiming to facilitate a broader trade agreement. The move would
mark a significant shift from current restrictions targeting China’s access
to advanced semiconductors used in AI and defense. While Trump publicly
2025

maintains a tough stance on Beijing, insiders suggest his campaign sees
relaxed export policies as leverage for economic negotiations. The
proposal has sparked debate over national security, technological
dominance, and U.S. trade strategy in the AI era.
5.23
China Proposes
Global AI
Cooperation
Organization at
Shanghai Summit
At the 2025 Shanghai Cooperation Organization summit, China
proposed a new global AI cooperation body aimed at improving
governance, risk mitigation, and joint development of AI technologies. The
initiative emphasizes equal representation and non-discriminatory access
to AI benefits, countering what China views as Western-dominated
frameworks. Beijing’s pitch includes collaborative R&D, shared safety
standards, and open infrastructure access. The proposal is part of China’s
broader diplomatic push to shape international AI norms and challenge
U.S.-led regulatory influence. Reactions are mixed, with some nations
expressing interest and others citing concerns over transparency.
By Retuers 🔗 July 26,
2025
5.24
U.S. White House
Releases AI
Playbook to Secure
Global Leadership in
AI
The White House has unveiled a comprehensive AI Playbook outlining
America’s strategy to lead the global AI race. It emphasizes responsible
development, national security, workforce readiness, and global
cooperation. Key elements include boosting federal AI R&D funding,
promoting ethical AI practices, accelerating AI talent pipelines, and
ensuring fair competition. The playbook also underscores public-private
partnerships and regulatory clarity to stimulate innovation while
safeguarding rights and safety. This initiative positions the U.S. to compete
with global AI powers like China and the EU.
By Asif Razzaq 🔗 July 27,
2025
5.25
Sam Altman
Cautions Against
OpenAI CEO Sam Altman warned that ChatGPT does not provide legally
confidential therapeutic interactions. Despite some users turning to it for
2025

Using ChatGPT as a
Substitute for
Therapy
mental health support, ChatGPT sessions are not protected under doctor-
patient privilege or HIPAA, making shared data potentially accessible.
Altman’s comments reflect broader ethical and regulatory concerns
around AI used in sensitive contexts like mental health. He emphasized
the need for user awareness and reinforced that ChatGPT should not
replace licensed mental health professionals, especially in serious or
emergency situations.
5.26
Anthropic Faces
Backlash Over
Claude API Rate
Limits Amid Soaring
Demand
Anthropic is under fire from developers after quietly imposing strict rate
limits on Claude API usage, especially affecting popular Claude 3.5
Sonnet. Users report downgrades from 1,000 to 100 requests/min,
severely hampering real-time applications. While Anthropic cites demand
surges and model availability constraints, developers criticize the lack of
transparency and communication. The incident raises broader concerns
about platform reliability and the trade-off between scaling access and
maintaining performance in commercial AI services.
2025
5.27
CBA Cuts 45 Jobs
Amid AI Automation,
Faces Union
Pushback
Australia’s largest bank, Commonwealth Bank of Australia (CBA), has laid
off 45 employees due to increased AI-driven automation, triggering sharp
criticism from the Financial Sector Union. The layoffs affect customer
service and operational roles, with the union arguing CBA is prioritizing
cost-cutting over job security despite record profits. CBA claims it’s
aligning workforce capabilities with emerging technologies, a trend
mirrored across the banking sector. The incident underscores growing
tensions between labor groups and corporations over the ethical
deployment of AI.
2025

6.1 Ai4 2025
Ai4 2025, held August 11–13 at 3799 S Las Vegas Blvd, brings together the
global AI ecosystem for its flagship annual conference aiml.events. Since
2018, Ai4 has become a cornerstone for both technical and business leaders—
including enterprise executives, AI startups, investors, policymakers, and
media. The event spotlights generative AI, AI agents, and practical applications
across industries, offering expert-led sessions, networking opportunities, and
hands-on workshops. Attendees gain insights into responsible human–
machine collaboration, emerging trends, and best practices. It’s essential for
anyone shaping the future of enterprise AI, from strategy and deployment to
ethics and innovation.
By AI & ML
Events 🔗 August 11 - 13,
2025
6.2
Pie & AI: Accra -
Future Forward:
Careers in the Age
of AI & Automation
Pie & AI: Accra brings together students, educators, and professionals on
July 26, 2025, at 10:00 AM WAT, to explore how AI and automation are
reshaping careers and education choices. This free event focuses on preparing
participants for a future where machine learning and intelligent systems
become integral to work and learning. Through expert talks, interactive
workshops, and networking, attendees will learn to identify emerging job
opportunities, develop relevant skills, and navigate educational pathways. If
you're planning your career or guiding others in an AI-driven age, this event
helps chart an informed, forward-thinking path.
By Pie & AI
by
DeepLearnin
g.AI
community
🔗 July 26, 2025
6.3
The 63rd Annual
Meeting of the
Association for
Computational
Linguistics
ACL 2025, the 63rd Annual Meeting of the Association for Computational
Linguistics, will be held from July 27 to August 1, 2025, in Vienna, Austria. It is
a leading conference in Natural Language Processing, gathering researchers,
academics, and industry experts. The event includes registration and opening
on July 27, main conference sessions from July 28-30, and workshops on July
31 and August 1. Topics cover machine translation, language modeling, ethics,
human-centered NLP, and more. Dr. Luke Zettlemoyer will deliver the opening
By ACL 2025
Vienna 🔗 July 27, 2025

keynote on efficient training of large language models. Sessions will be
accessible both in-person and online.
6.4
World Artificial
Intelligence
Conference (WAIC)
2025
WAIC 2025 is set to convene in Shanghai from July 26 to 28 (exhibition
continues through July 29) at the Shanghai World Expo Center. Under the
theme “Global Solidarity in the AI Era”, it brings together leading figures in AI
research, industry, investment, and governance. For the first time, it will feature
separate forums on breakthrough technologies, emerging industries,
humanities, and future trends. The conference includes extensive networking
opportunities and live streaming partnerships, enabling international
collaboration across universities, tech firms, and associations. Organized by
Donghao Lansheng Group, WAIC aims to promote global dialogue, innovation,
and cooperation in artificial intelligence.
By Donghao
Lansheng
(Group)
Co.,Ltd.
🔗 July 26-29,
2025
Conclusion
• July 2025 closes with AI gaining unprecedented momentum across innovation, business, and policy.
• AI is shifting from emerging tech to essential infrastructure with wide industry impact.
• Open-source models rival proprietary ones, with NVIDIA enabling LLM training on single GPUs and Hugging Face launching a unified CLI.
• Infrastructure demands are reshaping cloud strategies, with deals like OpenAI-Oracle, Armada’s portable data centers, and Intel's investment shift.
• Enterprise AI adoption is delivering real value, as seen in Citi's research, GitHub’s Copilot Agent Mode, and LegalOn’s $50M raise.
• Browsers are becoming AI agents—Edge adds Copilot Mode, Google tests Web Guide, and AI web referrals surged 357% year-over-year.

• AI safety is a growing focus, with efforts like Anthropic’s auditors, Holistic AI’s Grok-4 analysis, and warnings on ChatGPT misuse.
• Environmental and ethical concerns rise, shown by Mistral’s emissions data and impacts on programs like bug bounties.
• Competing AI governance visions from the U.S., Trump campaign, and China reflect the growing geopolitical stakes.
• Research in reasoning, efficiency, and multimodality shows AI's potential is still in early stages, with rapid progress ahead.

NewMind AI Weekly Chronicles - July'25 - Week IV

More Related Content

Similar to NewMind AI Weekly Chronicles - July'25 - Week IV (20)

Recently uploaded (20)

NewMind AI Weekly Chronicles - July'25 - Week IV