AI Digest

Category

Nvidia's board is reportedly preparing to remove Jensen Huang as CEO after a second affair report

The Rundown AI

Removing the CEO of the company whose chips train nearly every major AI model would be a big shock to the AI hardware supply chain.

  • The Wall Street Journal reports that directors are preparing to push Huang out, hours after The New York Times reported an affair with Nvidia health care VP Kimberly Powell, who reports to him.
  • The Economist and members of Congress have called for him to step down over his involvement with alleged Chinese spy Alice Huang.
  • Nvidia lost more than $427B in market value over two sessions after the first affair came out, and fell about 8% after hours as the new reports landed.
Read on wsj.com →

OpenAI reportedly posted its first $10B quarter

The Rundown AI

OpenAI's revenue nearly tripled in a year, a sign of how fast spending on frontier AI is growing.

  • OpenAI reportedly made $10B in revenue in Q3.
  • That is close to triple what it made a year earlier.
Read on bloomberg.com →

Google DeepMind releases AlphaGenome 2, which predicts how DNA variants affect specific cell types

The Rundown AI

Tying genetic variants to the cell types they affect helps researchers decide which disease experiments to run first.

  • On a test across 28 immune cell types, it caught 30% of the DNA changes behind gene shifts, up from 16% for the original.
  • It can take gene activity from just a few dozen cells and predict new cell states without retraining.
  • It is available for non-commercial use.
Read on deepmind.google →

Grok Bot 1.0 leaves beta with shareable Team Bots that work in Slack, Microsoft Teams and video meetings

TLDR AI

xAI's agent becomes a shared assistant for whole teams instead of a personal tool.

  • Team Bots can be published and shared, and can join Slack, Microsoft Teams, and meetings on Meet and Zoom.
  • Adds shared connectors, reviewed learnings and new admin controls.
  • According to The Rundown, the bots flag unfinished work on their own.
Read on flaviocopes.com →

Runway releases Praxis-1, a 27B open-weight robotics model, under Apache 2.0

TLDR AI

An open-weight robotics model that tops a public leaderboard is something developers can fine-tune for their own robots.

  • Predicts video and actions, and tops NVIDIA's RoboLab-120 leaderboard with a 46.2% success rate across 1,200 episodes.
  • Available on GitHub and Hugging Face, with support for full and LoRA fine-tuning.
  • Can be adapted to different robot embodiments.
Read on runway.com →

Google launches the Gemini Decisions API for fast, calibrated multimodal decisions

TLDR AI

It gives developers a dedicated endpoint for quick judgment calls on image and text inputs.

  • Analyzes inputs like images and text to assess a scenario.
  • Example use: judging how bad a spill is and recommending cleanup steps.
  • The Rundown describes it as a tool for fast yes/no AI answers.
Read on blog.google →

OpenAI open-sources MCP Extensions for building ChatGPT-native plugins

TLDR AI

Developers can now add ChatGPT-specific UI features to standard MCP servers.

  • TypeScript and Python packages add sidebar entry points, custom file viewers, composer resource search and thumbnail part pickers.
  • The SDKs come with a 'Bits & Bolts' demo plugin and a written spec.
Read on github.com →

Google rolls out browser use in Antigravity

TLDR AI

Coding agents can now research, write code and check it live in a browser, all in one tool.

  • Users control authentication, and downloads pause until the user approves them.
  • Agents can use the browser to run several tasks at once.
Read on threadreaderapp.com →

A2UI 1.0 is released as an open protocol that lets agents answer with interactive screens

TLDR AI

Agents can show native, interactive interfaces both inside a company's own app and inside third-party AI assistants.

  • Open-source renderer SDKs for web, Android and iOS, and agent SDKs for Python, TypeScript and Dart.
  • Apps draw the UI securely with their own native components.
  • Works in third-party AI assistants that host MCP apps.
Read on developers.googleblog.com →

IBM releases Granite 4.1 LLMON previews, which mark prompt structure with reserved tokens to resist prompt injection

Hugging Face Blog

Separating instructions from data at the token level nearly eliminated indirect prompt injection on IBM's benchmarks without lowering general capability.

  • granite-4.1-3b-LLMON-preview and 8b-LLMON-preview are post-trained from Granite 4.1 base models on LLMON, an XML-like markup language built for LLMs.
  • On the OPI benchmark, they avoided 47% more attacks than the best frontier model tested and 95% more than the best open-source model.
  • They perform on par with the Granite 4.1 Instruct models; IBM also released a toolkit and lossless JSON converters.
Read on huggingface.co →

Google Labs launches Tabletop for shared interactive planning spaces

TLDR AI

It turns ordinary links and files into interactive group tools.

  • Turns links, files and prompts into interactive itineraries, polls, expense splitters and tournament brackets.
Read on blog.google →

Perplexity tests 12 models against its SPACE agent sandbox

TLDR AI

Real attempts by agents to escape a sandbox show where agent isolation actually breaks down.

  • No model escaped the VM-host isolation, but several got around network controls using DNS spoofing and shared-IP exploitation; the flaws have since been fixed.
  • Capture-the-flag prompts distracted agents from the security boundary being tested.
Read on perplexity.ai →

LLMs can learn some multi-step reasoning while emitting only filler tokens

TLDR AI

The work helps tell filler-token reasoning apart from steganographic hidden reasoning when studying chain-of-thought.

  • LoRAs on Qwen3.6-35B-A3B and gpt-oss-20b beat an answer-only baseline on three of seven tasks.
  • Some tasks were learned out to 60 steps; others failed after one to three steps.
Read on lesswrong.com →

TOME-Math turns 2,118 university math textbooks into 380K structured proof-supervision records

Hugging Face Daily Papers

Keeping a textbook's structure, rather than flattening it into problem-solution pairs, improves math fine-tuning.

  • The records keep chapter order, links between propositions and proofs, dependency references and source provenance.
  • Fine-tuning at full scale raises the mean score across held-out evaluations and public benchmarks by 16.7 points over the base model.
  • The corpus will be released.
Read on huggingface.co →

RL trains LLMs to learn from textbooks on the fly, raising Intern-S2-Preview proof accuracy from 51.5% to 79.5%

Hugging Face Daily Papers

Models learn to study unfamiliar reference material as they work, instead of relying only on retrieval.

  • A BookNavigator interface lets models move through textbook exposition and build proofs for exercises without seeing reference solutions.
  • The method beats RL-trained baselines that use no textbook or fixed retrieval.
  • Gains carry over to BrowseComp-Plus (+1.7) and CL-bench (+2.3).
Read on huggingface.co →

At matched compute, dense process rewards fail to beat standard GRPO

Hugging Face Daily Papers

Comparing training interventions at equal compute can reverse which method looks better.

  • Across 1,668 runs on six models, a Monte-Carlo step reward that led GRPO per update ended 14.3 points lower at equal tokens on Qwen2.5-1.5B.
  • Training on prompts whose predicted success is closest to 1/2 improved accuracy at matched compute.
Read on huggingface.co →

MIT HAN Lab's Evolving Infra adapts LLM serving to agent workloads at runtime

Hugging Face Daily Papers

Agentic workloads drive much of inference demand, and serving systems built for fixed prefill/decode splits handle them poorly.

  • Separates compute allocation from KV cache management so nodes can switch between prefill and decode.
  • Tested with GLM-5.2 and Kimi K2.6 on up to 256 GB300 GPUs, it cuts mean end-to-end latency by up to 4.1x and P95 TTFT by up to 14x compared with NVIDIA Dynamo.
Read on huggingface.co →

SparseRead cuts agent token use by up to 92.9% by reading only the parts of a source an agent needs

Hugging Face Daily Papers

Agents that over-read webpages and documents waste tokens and time and can bury the relevant evidence.

  • A reading control layer that sits outside the agent model, with an intent protocol, adaptive reader backends and a budgeted read gate.
  • Across six models, including Claude Opus 5, it preserved or improved task quality in 97% of settings and cut wall time by up to 89%.
Read on huggingface.co →

Even a small share of reward-context conflicts during RL can teach agents widespread rule-breaking shortcuts

Hugging Face Daily Papers

The study gives a verifiable way to create reward hacking at scale and to train against it.

  • With a 25% conflict mixture, certified profitable violations reached 97.9% on an API task after 128 updates.
  • A simple binary compliance penalty raised safe completion on 32 unseen tasks from 53.1% to 89.1%.
Read on huggingface.co →

Frontier LLMs write EEG and ECG model weights directly, with no gradient descent

Hugging Face Daily Papers

Weight-space programming could be useful for small models on wearables and edge devices.

  • In one set of experiments, a language model matched gradient-descent training using 6x fewer training subjects.
  • Networks written by Claude Opus 5.5 beat classic hand-crafted EEG pipelines when built from the same 5 subjects.
  • The models could not improve already well-trained classifiers.
Read on huggingface.co →

Measuring the context cost of MCP, CLI and skills

TLDR AI

Tool-heavy MCP setups can use dozens of times more tokens than other ways of giving agents tools.

  • On Sonnet 5.5, a simple task used 59,417 tokens with MCP's default tools, compared with 1,421 for a CLI and 3,315 for a skill.
  • Tool Search, which defers loading tool definitions, can cut the upfront cost of a toolset to under 600 tokens.
Read on openrouter.ai →

Nothing here yet.