OpenAI developed a new way to train small AI models with internal mechanisms that are easier for humans to understand.

Language models like the ones behind ChatGPT have complex, sometimes surprising structures, and we don’t yet fully understand how they work.

In a new research, team train “sparse” models—with fewer, simpler connections between neurons—to see whether their computations become easier to understand.

Efficient Self-Improving Agent Systems. AgentEvolver lets AI agents improve themselves instead of requiring manual prompt tuning.

They use three core mechanisms: self-questioning, self-navigating, and self-attributing.

Agents evaluate their own work, spot failures, and write better instructions for themselves.

This leads to a self-improvement loop capable of running without human oversight.

Shows better performance across benchmarks with less manual work.

The framework works by having agents evaluate their own performance on tasks, identify where they failed or underperformed, and then generate improved behavioral instructions for the next iteration.

The results are impressive.

Agents using this approach show measurable performance gains across diverse benchmarks compared to static configurations, all while reducing the overhead of constant manual optimization.

Google is working on multi-agent systems to help you refine ideas with tournament-like evaluation.

Each run takes around 40 minutes and brings you 100 detailed ideas on a given research topic.

2 new multi-agents are being developed for Gemini Enterprise:
- Idea Generation - "Create a multi-agent innovation session"
- Co-Scientist - "Drive novel scientific discovery with Co-Scientist"

Co-Scientist 3-step workflow 👀
- Tell Co-Scientist what you plan to research, point it to relevant data, and set your evaluation criteria.
- A team of agents will generate ideas on your topic using their available data
- The agents will evaluate the ideas against your criteria and rank them, tournament-style

Google is not only automating research but also preparing a product that will enable others to do so.

Android creator Andy Rubin is launching a new humanoid robotics startup, "Genki Robotics," in Tokyo.

The company is operating in stealth mode, tapping Japan's engineering talent to enter an already crowded field.

During his tenure at Google, Rubin spearheaded an ambitious robotics division, leading the acquisition of numerous startups in 2013, including the high-profile Japanese humanoid firm Shaft, a spin-off from the University of Tokyo.

His interest in legged locomotion, a core challenge in humanoid development, is well-documented. At a 2018 tech conference, Rubin, then leading the incubator Playground Global, predicted a future of "legs everywhere." He argued that legged systems are essential for navigating human-centric environments, such as climbing stairs or using elevators for "last-mile delivery"—tasks impossible for wheeled machines.

MIT and Oxford released their $2,500 agentic AI curriculum on GitHub at no cost.

15,000 people already paid for it.

It covers patterns, orchestration, memory, coordination, and deployment.
A strong roadmap to production ready systems.

Google DeepMind introduced WeatherNext 2 is most advanced system yet, able to generate more accurate and higher-resolution global forecasts.

The model’s improved performance is enabled by a new approach called a Functional Generative Network, which can generate the full range of possible forecasts in a single step.

Team added targeted randomness directly into the architecture, allowing it to explore a wide range of sensible weather scenarios.

MIT Introduced JiT (Just image Transformers)

JiTs are simple large-patch Transformers that operate on raw pixels, no tokenizer, pre-training, or extra losses needed.

By predicting clean data on the natural-data manifold, JiT excels in high-dimensional spaces where traditional noise-predicting models can fail.

On ImageNet (256 & 512), JiT achieves competitive generative performance, showing that sometimes going back to basics is the key.

GitHub.

Physical intelligence introduced a new model π*0.6

π*0.6 can more than double throughput over a base model trained without RL, and can perform real-world tasks: making espresso drinks, folding diverse laundry, and assembling boxes.

Team trained a general-purpose value function on all of own data, which tells the π*0.6 VLA which actions are good or bad. By asking π*0.6 to produce only good actions, researchers get better performance. Team call this method Recap.

π*0.6 can then collect more autonomous data, which can be used to further train the value function and further improve π*0.6.

During autonomous data collection, a teleoperator can also intervene and provide corrections for significant mistakes, coaching π*0.6 further.

Quantitatively, training π*0.6 with RL can more than double throughput (number of successful task executions per hour) on the hardest tasks and cut the number of failures by as much as a factor of two.

Google DeepMind just released Gemini 3 that helps you learn, build and plan anything.

It comes with state-of-the-art reasoning capabilities, world-leading multimodal understanding, and enables new agentic coding experiences.

PyTorch creator Soumith Chintala has joined Thinking Machines Lab.

Official exit from Meta: Nov 17. New gig at TML: Nov 18.

He says the people there are "incredible" and he is already back to "building new things." The AI talent war continues.

Can LLMs really behave like human investors? How do micro-level behaviors drive macro-level market dynamics?

TwinMarket offers an answer by placing thousands of LLM-driven investors in a realistic stock market environment that incorporates social networks, news, and behavioral biases.

This setup lets us watch bubbles, crashes, and herding emerge from individual decisions.

Calibrated on real market data and grounded in behavioral finance, TwinMarket scales to 1,000+ agents, reproduces key stylized market facts (volatility clustering, fat tails, etc.), and reveals how social interaction and cognitive biases jointly drive systemic risk.

The work is accepted to NeurIPS 2025 and received the Best Paper Award at the ICLR 2025 Financial AI Workshop.

GitHub.

Meta introduced a new generation of Segment Anything Models:

1. SAM 3 enables detecting, segmenting and tracking of objects across images and videos, now with short text phrases and exemplar prompts.

2. SAM 3D brings the model collection into the 3rd dimension to enable precise reconstruction of 3D objects and people from a single 2D image.

Elon Musk’s xAI Introduced Grok 4.1 Fast and the xAI Agent Tools API.

With a 2M context window, it shines in real-world use cases like customer support and deep research.

#DeepSeek just released LPLB

Linear-Programming-Based Load Balancer (LPLB) is a parallel load balancer that leverages linear programming to optimize expert parallel workload distribution for MoE (Mixture-of-Experts) models.

Kimi dropped Seer: Online Context Learning for Fast Synchronous LLM Reinforcement Learning.

Ethereum co-founder Vitalik Buterin warned that if BlackRock and other large institutions keep expanding their ETH holdings, Ethereum faces two risks:

1) decentralization-minded builders could be crowded out, weakening the community;

2) base-layer choices optimized for institutions (e.g., ~150 ms block times) could make it infeasible for typical users to run nodes, driving geographic/network centralization.

Nabla announced JAM-2 — the first AI model capable of generating drug-quality antibodies straight from the computer, with industry-leading success rates.

> Drug-like affinities: Picomolar to single-digit nanomolar antibody binders for half of 26 targets while testing <45 designs each.

> Unlocking hard targets: Up to 11% success rate for direct on-cell GPCR binders; top antibody hits in the single-digit nanomolar range.

> Unprecedented epitope breadth: JAM-2 routinely designed antibodies that hit 30–70% of user-defined epitopes, now enabling intentional design of biology — not chance discovery.

> Drug-like developability: Over 50% of antibody designs passed core industry developability criteria with zero optimization.

> Massive leverage: A four-person team prosecuted 16 targets in parallel in < 1 month.

JAM-2 is the first de novo antibody design capability ready for front-line use in drug discovery, matching or surpassing traditional discovery approaches.

Physical Intelligence has raised $600m at a $5.6 bil valuation, w/ CapitalG leading

Earlier this week, Pi released a new RL model that it says can help robots do tasks like fold laundry, pack boxes more quickly

Ai2 presented Olmo 3, a fully open LM suite built for reasoning, chat, tool use, and an open model flow—not just the final weights, but the entire training journey.

At the center is Olmo 3-Think (32B)—a fully open 32B-scale reasoning model.

Olmo 3 opens the model flow – pretraining, mid-training, & post-training – plus data recipes & code so you can see how capabilities are built + customize any stage.

Meet the Olmo 3 family:
1. Olmo 3-Base (7B, 32B)—foundations for post-training with strong code, math, & reading comprehension skills
2. Olmo 3-Instruct (7B)—multi-turn chat + tool use
3. Olmo 3-Think (7B, 32B)—“thinking” models that show their reasoning.

All designed to run on hardware from laptops to research clusters.

Sakana AI co-founded by Llion Jones (one of the original 8 inventors of Transformers at Google), now argues it's time to move past them.

Their NeurIPS 2025 spotlight paper, “Continuous Thought Machines”, outlines a possible next step.

Anthropic is working on a new Skill creation flow where Claude itself can create the Skill for you.

Other things to expect this week:
- Opus 4.5 is rumoured to be launched today (not confirmed).
- Claude code desktop app?
- New referral program
- Megabrain?