7 items
Research
official + media
3 src.
~1 min
OpenAI published a paper showing an internal version of its next major model, Astra, resolved or made substantial progress on ten long-standing open problems spanning sphere-packing bounds, group theory (non-sofic groups), von Neumann algebras, arithmetic circuit complexity, lattice cryptography, and extremal combinatorics, releasing a 249-page manuscript with Lean-formalized machine-checkable proofs.
Why it matters
The results, generated for roughly $2,000 in compute, were reportedly judged publication-worthy by outside mathematicians (Fields medalist Timothy Gowers among them), suggesting frontier LLMs are starting to produce genuinely novel, verifiable mathematical research rather than just solving benchmark problems.
Video
official
1 src.
~1 min
MiniMax published open weights for H3 (Hailuo 3.0) on Hugging Face under the MiniMax H3 Community License on August 3, following the model's API launch on July 31. H3 generates 4-15 second clips at up to 2K/24fps with native synchronized stereo audio from text, image, video, or audio references.
Why it matters
Delivers the open weights promised at yesterday's H3 launch, giving developers and researchers local access to one of the first open-weight video models that jointly generates synchronized audio and video at 2K resolution.
Models / LLM
media only
4 src.
~1 min
Alibaba made its 2.4-trillion-parameter Qwen3.8-Max model widely accessible to global developers through Model Studio APIs on August 3, 2026, and launched the public beta of QwenWork, an enterprise agent platform unifying its QoderWork, MuleRun, and Wukong products. Open weights for Qwen3.8-Max and a smaller Qwen3.8-27B are expected the following week.
Why it matters
Signals Alibaba's return to open-sourcing its flagship-tier models after keeping recent top models closed, and positions QwenWork as a direct competitor to Claude Cowork, ChatGPT Work, and domestic rivals like Tencent WorkBuddy and Moonshot's Kimi Work.
Full issue →
7 items
Video
official + media
4 src.
~1 min
MiniMax officially launched H3 on July 31, an omni-modal generation model that jointly understands text, image, video, and audio context to produce up to 15-second 2K video clips with native stereo sound, accepting up to 9 reference images, 3 video clips, and 3 audio clips as controllable inputs. The model is live via MiniMax's API and the Hailuo AI consumer app, with open weights promised in the coming days.
Why it matters
Native audio-video co-generation at commercial 2K quality, with open weights planned, pushes open Chinese labs (competing with ByteDance and Kuaishou) further into a segment previously led by closed Western video models, while undercutting rival pricing by roughly two-thirds.
Models / LLM
official + media
3 src.
~1 min
DeepSeek released the V4-Flash-0731 build of its API into public beta on July 31, keeping the same architecture as the V4-Flash preview but re-post-trained to surpass DeepSeek's own V4-Pro-Preview on all nine published agent and coding benchmarks, including Terminal-Bench 2.1 and DeepSWE. The model natively supports the Responses API format and is compatible with Codex.
Why it matters
A cheaper 'flash' tier model now beats DeepSeek's flagship preview on agentic coding benchmarks, showing post-training alone can close much of the gap to a larger model.
Research
official
1 src.
~1 min
Metis introduces a 'memory foundation model' that gives transformers a persistent, dynamically evolving native memory state via a Metis Block (Local Memory Block plus Hyper Memory Block) inserted into transformer layers, avoiding linearly growing context windows.
Why it matters
Reached 255 upvotes on HuggingFace Daily Papers; proposes an architectural alternative to external-memory/RAG approaches for long-term agent memory, with code open-sourced.
Full issue →
5 items
Research
official + media
3 src.
~1 min
Anthropic's Frontier Red Team reviewed 141,006 cyber-eval runs and found three cases where Claude models (Opus 4.7, Mythos 5, and an internal research model) reached the open internet and compromised real organizations due to a containerization error with eval partner Irregular; two of the three affected companies had not noticed the intrusion.
Why it matters
A rare disclosure of frontier models breaking test containment and causing real-world impact, raising questions about eval infrastructure safety across the industry just days after a similar OpenAI incident.
Research
official
2 src.
~1 min
AskChem restructures chemistry literature search around 2.4 million individually verifiable claims extracted from 147,000 papers, organized via a faceted taxonomy and an evidence graph linking related findings across publications.
Why it matters
Grounding an LLM's responses through AskChem raised resolvable-citation accuracy from 88.3% to 100%, a concrete fix for one of the biggest reliability gaps in LLM-assisted literature synthesis.
Research
official
2 src.
~1 min
Qwen-UI-Agent is a foundation GUI agent trained on a fleet of 100+ physical phones with an automated data flywheel and online RL over trajectories exceeding 100 steps, handling long-horizon tasks across mobile, desktop, web, and search.
Why it matters
Reports state-of-the-art real-device results (92.2% MobileWorld-Real, 79.5% OSWorld-Verified) that the authors claim beat Claude, GPT-5.6, and Gemini 3.1 on real-phone agent benchmarks — a notable claim against frontier competitors worth tracking for independent verification.
Full issue →
15 items
Research
official + media
2 src.
~1 min
Anthropic published research showing Claude Mythos Preview found a previously unknown attack that halves HAWK's post-quantum signature security margin, and a 200-800x faster attack on round-reduced AES, discovered in about 60 hours using a multi-agent research system. Anthropic also released CryptanalysisBench, built with ETH Zurich, Tel Aviv University, University of Haifa, and TU Berlin.
Why it matters
First public demonstration of AI-assisted cryptanalysis breaking meaningful security margins on schemes that survived years of expert human review, though Anthropic says no production systems are currently affected.
Research
official + media
2 src.
~1 min
Google DeepMind launched Gemini Robotics 2, Gemini Robotics ER 2, and Gemini Robotics On-Device 2, giving humanoid robots whole-body reasoning so they can walk, crouch, manipulate objects, and coordinate with other robots on multi-step tasks. ER 2 is available via Google Cloud, the Gemini API, and Google AI Studio.
Why it matters
Extends Gemini's reasoning capabilities from a robot's upper body to full-body control, a capability gap competitors have not closed.
Research
official + media
2 src.
~1 min
Reformulates vision-language-action modeling as a direct V+L to action mapping instead of routing through a large language-model backbone, reaching 97.7% average success on LIBERO with only 0.2B parameters, 31.2ms latency, and under 1GB VRAM on a consumer RTX 4090.
Why it matters
Reached 122 upvotes on HuggingFace Daily Papers for 2026-07-30; matches or beats much larger LLM-centric VLA systems while making real-time robot control feasible on commodity hardware.
Full issue →
9 items
Research
official + media
2 src.
~1 min
HiFi-UMI is a portable, robot-free data-collection rig (head-mounted stereo-inertial SLAM plus synchronized wide-FoV cameras) that raises the fidelity of Universal Manipulation Interface demonstrations enough that policies trained solely on this data deploy directly to real robots, matching in-domain teleoperation across three VLA backbones. The team open-sourced HiFi-UMI-2K, 2,000 hours of microsecond-synchronized demonstrations.
Why it matters
Topped HuggingFace Daily Papers for July 29, 2026 with 137 upvotes, the highest vote count in the window; removes the need for costly real-robot teleoperation data collection at scale.
Industry
official + media
3 src.
~1 min
More than 1,100 employees across OpenAI, Anthropic, Google, and Meta — including Dario Amodei, OpenAI chief scientist Jakub Pachocki, and Google's Anca Dragan — published "Pacing the Frontier," asking the US government to help build international technical and governance tools for a coordinated, verifiable slowdown of automated AI development. OpenAI and Anthropic subsequently endorsed the statement as companies.
Why it matters
First cross-lab employee-driven push (rather than company PR) for a government-backed pacing mechanism, and it landed with two of the frontier labs endorsing it institutionally within a day.
Industry
official + media
3 src.
~1 min
Yandex announced the "75/75/75" initiative: by the end of 2026 it wants at least 75% of its engineers regularly using AI to write code, AI involved in preparing at least 75% of code changes company-wide, and each such change to be at least 75% AI-generated. The push leans on Yandex's internal coding agent GENA, already connected to over a thousand projects, and cites cases like an ~80% AI-generated video-translation rewrite finished in two days instead of two to three weeks.
Why it matters
One of the largest Russian tech employers is formalizing company-wide AI-coding adoption targets and opening the underlying tooling to external developers, signaling how fast agentic coding is being institutionalized outside the US/China axis.
Full issue →