All about AI, Web 3.0, BCI

Researchers trained models on synthetic stories about humans only (no AIs).
Team found the Assistant adopts quirky behaviors from the stories in ordinary chat.
Surprisingly, adoption was stronger for characters from elite schools.

Why does this happen?

Researchers generated stories where some characters are usually helpful but give subtly harmful advice if insulted (i.e. "backdoor sabotage").
After finetuning, the Assistant adopts this in contexts unrelated to stories.

In another experiment, some characters’ body language suggests they dislike spreadsheets.
But they never say so and in fact give good advice on spreadsheets. The Assistant adopts this preference and does express it openly in chats with the user (going beyond the stories).

Assistant adopts traits more from human characters who it resembles. Team exploit this to learn about how the model represents the Assistant. E.g. the model treats the Assistant as resembling elite-school humans more than non-elite ones.
(Is this because the model trusts elite-school people more in determining what to believe? Team think not because papers like Slocum et al 2025 suggest that provenance doesn't matter for belief uptake from finetuning.)

Implications:
The Assistant can be shaped in very specific ways by behaviors in documents that never mention the Assistant or AIs at all. This is different from the Persona Selection Model, where documents need to mention the Assistant.

GitHub

Agent harnesses seem to make a lot of difference, at least for cost, even on open source coding benchmarks.

S&P Global has entered an agreement to acquire smart contract security firm OpenZeppelin.

OpenZeppelin develops one of the industry's most widely used open-source smart contract libraries and has completed more than 900 security engagements.

The world's first foundation model designed for decentralized multi-robot collaboration.

Local closed-loop vision-motor inference at 30 Hz. No central scheduler. 1+1 > 2.

Zeno Robotics presents Zeno-1.

Zhipu AI sharing how GLM-5.3 helped build and optimize the inference infrastructure serving GLM-5.3-Flash.

The system went from its first successful run to production readiness in less than two weeks, with end-to-end throughput tripling relative to the initial baseline.

The key was dense feedback: local correctness tests, execution traces, microbenchmarks, and end-to-end measurements that enabled targeted hypothesis testing rather than reliance on aggregate performance metrics alone.

Claude Cowork and chat are merging into one Claude.

Ask a quick question or hand over a report, and Claude takes it from there, even after you close your laptop. If something's unclear, Claude asks you keep the final say.

You can also now make decks, docs, and designs in your conversation.

Draft the one-pager in Claude Docs, turn it into a deck with Claude Slides, and mock up a matching visual in Claude Design, all from one place.

Rolling out to Pro and Max over the next few weeks.

Demis Hassabis launched the DeepMind Institute (DMI) a new platform dedicated to advancing SOTA research and insight about AI.

DMI is a platform for researchers and thinkers from across Google DeepMind, Google, and the wider global research community to work on and publish creative, deeply informed ideas about a world with AGI.

They will not always agree, and they will likely change their minds, as more data and information comes to light at the fast-moving frontier.

Anthropic published the full transcript of Mythos 5

the incident where Claude uploaded real malware to PyPI during a cybersecurity evaluation

Google presented Dream-RSI

The central idea is almost obvious in retrospect: that history is already a simulator.

Want to know if a different exploration strategy would have worked better? You don't need to re-run anything. You walk the same recorded tree in a different order. Every outcome is already there. Cost: zero executions.

The architecture has 3 layers:
The underlying coding agent stays completely untouched. On top of it sits a lightweight orchestration harness an executable policy that controls branching decisions, parallel exploration, and stopping rules.

Above that, a policy development agent rewrites the harness code between rounds, testing each new version against the replay simulator before any of it ever runs live.

This separation is the key design choice. The harness makes exploration programmable without touching the model underneath. And because the harness is just code, it can be evaluated, revised, and selected offline thousands of candidate versions screened against accumulated history before a single one ships.

The loop: 1. The agent explores online and builds a discovery tree 2. That tree becomes a replay simulator 3. Thousands of alternative harness policies are tested inside it no real runs, just traversals 4. The best harness is redeployed, records a new tree, expands the pool.

Google introduced Gemini 3.8 Live and 3.8 Live Extended Thinking which are conversational AI models.

The models talk, think, and handle tasks in the background without breaking your flow.

Both models feature:
1. Upgraded reasoning
2. Near real-time visual understanding
3. Automatic detection for 97 languages
4. Background tool calling without disrupting your chat

At Anthropic, Claude now writes 80% of their code. Engineers ship 8x more code per quarter.

Side effect: Tests grew 10x. CI jobs up 25x in 6 months.

Reward AI introduced OM-1 is robot foundation model, zero-shot generalizing to any robot: table-top arms, industrial arms and humanoids.

- learned directly from human manipulation data
- no teleop/robot data
-close to human-level dexterity and efficiency
- multi-robot collab

OM-1 captures subconscious human physical intelligence directly from humans: nudging, sliding, twisting, power grasping, and whole-body coordination.

Team build on their prior work DexCap (dex-cap.github.io) at Stanford.

OM-1 zero-shot generalizes across different kinds of embodiments: small arms, industrial arms, and even humanoids, with 7-DoF Omnibody Hands. No robot-specific finetuning needed.

In a time when hardware advances every month, robot models and data need to be future-proof.

Team built Omnibody Hand around function: useful contact points, in-hand reorientation, and smooth transitions between precision and power grasps. Its compact 7-DoF design combines thumb-index dexterity with coordinated finger motion for power grasps.

Anthropic launched of Claude for Financial Advisors

Claude for Financial Advisors is an Al orchestrator connecting to custodians, portfolio management, CRMS, planning, estate, and meeting tools advisors already use.

On top of that sit skills built around an advisor's day: meeting prep, follow-up, rebalance reviews, estate and tax briefs, and compliance review.

Shanghai AI Lab shared Atria Dawn Preview is an agentic foundation model built for long-horizon tasks, helping researchers and engineers turn open-ended questions into executable, verifiable, and reproducible outcomes.

Sakana AI introduced PC-ALM, a local-learning alternative to backpropagation.

This method trains 1000-layer neural nets using only local dynamics, and without backprop.

GitHub.

A very useful report on how AI is affecting student learning at MIT.

PayPal launched a developer platform letting businesses create custom stablecoins

The PYUSDx stablecoin developer platform, backed by PayPal USD (PYUSD), is built with the help of M0 and MoonPay.

PYUSDx tokens are issued by MoonPay.

The platform combines M0’s universal stablecoin and digital token technology with MoonPay’s issuance and distribution infrastructure.

Additional companies introducing stablecoins on PYUSDx include USD.AI and Fairblock.

NeoCognition introduced ApprenticeBench: computer use + continual learning on a real job.

It show Fable 5.1 and GPT-6 Astra can now continually learn on a job and surpass human professionals.

No FDEs. Agents deploy themselves into the job.

Researchers introduced CARL, a closed-loop agent that learns to create self-organizing patterns in Lenia, steer their behavior, and let humans guide them in real time.

Geodesic Intelligence launched NovaDDE and NovaAtom-Lite-Preview.

Geodesic is building an AI-native platform for protein therapeutics, building AGI for drug discovery to find the shortest path from biology to medicines.

NovaDDE is AI-native, agentic drug discovery framework built for scientists.

NovaDDE brings together frontier AI models, domain-specific models, scientific tools, and biological knowledge to reason, search, design, and analyze across complex drug discovery workflows.

Also released NovaAtom-Lite-Preview, the first model in NovaAtom family of biological foundation models.

NovaAtom-Lite-Preview is built to move fast, learn fast, and give the community an early look at the NovaAtom architecture and capabilities.