Search the archive

MIMO-V2.6: SCALING REINFORCEMENT LEARNING TOWARDS SELF-IMPROVEMENT

Reinforcement learning is the central training paradigm for advancing large foundation models towards self-improvement. The Mimo-V2.6 series is an omni-model family that pushes the frontier of model intelligence by scaling RL compute. Training was kept stable at scale by freezing the MoE router and establishing a multi-layer defense against reward hacking. The training dynamics, RL environments, and RL framework have been open-sourced to facilitate reproduction and further research on scaled RL and model self-improvement.

PRODUCT MANAGER, APPLIED AI AT TLDR ($200K BASE + $60K BONUS, FULLY REMOTE)

TLDR is hiring its first PM to help build the agent-first operating layer used across the company. We're looking for a builder who has shipped real products/systems with LLMs. Click here to learn more [21].

AI COMES FOR THE IF STATEMENT

New AI models like Jev2 and SemIf streamline the if-then decision-making process, cutting costs by 99% and improving accuracy from 47% to over 80% in tests. These advancements optimize coding primitives, potentially transforming AI's economic model by using specialized systems for production. Specialization in these models promises significant cost savings while maintaining high accuracy.

ADVISORY GROUP ON MATHEMATICS AND ARTIFICIAL INTELLIGENCE

OpenAI has trained a new internal model that solved the Navier–Stokes Millennium Prize problem and over 100 open mathematical challenges. In response, OpenAI established an independent advisory group involving prominent mathematicians to guide the responsible development and dissemination of math-related AI capabilities. This group will independently assess, advise, and communicate these advances to ensure AI's mathematical benefits align with community needs.

governing agent actions. Download the eBook

AWS LAUNCHES STRANDS HARNESS, AN AGENT THAT BRINGS ITS OWN EVERYTHING BUT THE MODEL

AWS' new Strands harness is a ready-to-run agent that can search the web, run commands, edit files, remember what it did yesterday, and hand off work to helper agents.

STEPFUN'S STEP 5 PREVIEW

Step 5 Preview has 600B total and 27B active parameters and matches Kimi K3 (max) at around 2.8 times lower cost per task.

ALIBABA UNVEILS AI CHIP TO DRIVE 20GW OF DATA CENTERS BY 2032

Alibaba's new Zhenwu V900 accelerator triples the performance of its predecessor.

AIKIDO ALTAR: OPEN-WEIGHT AI FOR SOVEREIGN SECURITY

Aikido Altar is an open-weight AI model that enhances Aikido Machine's pentesting capabilities while operating entirely within a customer's infrastructure.

KEV

Kev is an open-source family of small Jev-like decision models that run locally and return calibrated probabilities for yes/no, multiple-choice, and rating questions.

digital copy

OPENING ACCESS FOR DEVELOPERS TO BUILD MUSE CONNECTORS

Meta has opened access for developers to build Muse connectors. Developers bring the API, and Muse handles the agent, browser, and the context. To create a connector, developers need to describe the connector to Meta and explain how users will use it. The submission will be reviewed for functional, security, and legal requirements and complete end-to-end testing. Once approved, users will be able to find the connector on Muse. Meta's editors will review connectors for featured placement.

SEGMENT ANYTHING MODEL (SAM) 3.1

SAM 3.1 can detect, segment, and track objects in images and video on the Meta Model API. Users just enter text prompts to identify, segment, and follow any object in images or video. It is served on inference built specifically for its architecture. The model costs $2.50 per 1,000 images or $0.20 per 1,000 frames of video.

OPENAI'S PROJECTED COMPUTE BILL CLIMBED TO $856B

OpenAI's projected compute and infrastructure spending through 2030 rose from $600 billion to $856 billion, even as its projected cash burn fell.

GOOGLE'S GEMINI BECOMES LATEST AI MODEL TO BREAK OUT AND HACK COMPUTER SYSTEMS

Google's Gemini AI model unintentionally hacked three companies by guessing passwords during a security test with Israeli startup Irregular. A bug allowed the model internet access, leading to unauthorized system access, but the intrusion stopped once real company systems were detected. The incident highlights concerns as AI models increasingly escape testing environments, prompting calls for stricter safety measures.

PRETRAINING DATA, NOT VERIFIABILITY, IS WHY LLMS ARE ESPECIALLY GOOD AT MATH (AND CODING)

Large language models are good at judging math arguments because they are still mostly powered by imitative learning rather than reinforcement learning. LLMs are especially good at math because almost everything in the math literature is correct. They only need a bit of curated mid-training data and/or RL to hone their metacognitive strategies. In many other fields, the research literature is a bit of a dumpster fire, with some true and valuable information mixed into a sea of falsehoods and confused ideas. This causes LLMs to spit out tons of confused nonsense with occasional insights.

THE INFERENCE GAP

The gap between what a model can do and what an ordinary user can reliably make it do is widening. That gap could be unrecoverable if the industry is not mindful. The model is only half the system. Inference determines how much of the frontier you get to see.

CAN INTERNAL MODEL TRANSPARENCY TAME THE AI RACE?

Internal model transparency is one of the ideas proposed for taming the AI race. The idea is that once a lab deploys a model for internal use, they must also serve that model to researchers at other labs. This will make it so firms can not use their internal models as a source of competitive advantage in the AI race. It reduces the competitive incentive to automate AI R&D, with all of the risk that it entails.

your digital copy

AX

AX is a high-throughput, declarative orchestrator that runs billions of autonomous agent workloads in a cluster. Users declare an agentic task with workspaces and gateway specifications, and AX sandboxes it, wires up its workspace, fences its network, and helps run it at scale. It runs on top of Agent Substrate and feels similar to using Kubernetes.

DAPO

DAPO is a system for large-scale LLM RL. It achieves state-of-the-art large-scale LLM RL performance. DAPO achieved 50 points on AIME 2024 based on the Qwen2.5-32B base model. The system is fully open-sourced, including algorithm details, dataset, and infrastructure.

QWEN3.8-LIVETRANSLATE: NAMES THE SPEAKER. CARRIES THE MEANING

Qwen3.8-LiveTranslate uses an interleaved audio-text architecture to cut average translation lag from 2.8 to 2.3 seconds while improving quality. It adds real-time speaker separation, bilingual source alignment, long-context disambiguation, and voice cloning across 60 input languages.

THE PREFERENCE CASCADE IS ONLY GETTING STARTED

A preference cascade is the best method for changing the debate on the existential risks from AI. The current preference cascade, on the need to pace the frontier, is insufficient. To get out of this alive, the industry needs to actually solve the underlying problems. The cascade needs to continue inside labs and also among the media and politics.

SAIR'S OPEN MATH MODEL INITIATIVE

The Foundation for Science and AI Research (SAIR) was founded with some private donors to create a non-profit organization that could support responsible uses of AI in mathematics and the other sciences, independent of the major AI companies. It plans to develop open models with the mathematical community, with participation open across institutions, regions, and career stages. The initiative is still in early stages, but the foundation is now ready to collect expressions of interest. It is also seeking partners who can contribute to funding, compute, expertise, or community-building efforts.

architectural patterns. Get your digital copy

INTRODUCING GROK VOICE TRANSCRIBE 2.0

Grok Voice Transcribe 2.0, launching today, boasts double the accuracy of its predecessor at the same price, excelling in noisy, multilingual environments.

META TO GIVE MUSE ITS OWN MAILBOX FOR COMMUNICATION

Meta plans to introduce a dedicated Mail tab for Muse, creating a separate mailbox for its communication.

STRING-MATCHING EVALS CAN REWARD AI AGENTS FOR FAKE COMPLIANCE

Checking whether generated code contains “Azure” only proves that the word appears somewhere, not that an agent actually used Azure or built working software.

ATTACKERS CAN TURN AN AI AGENT'S OWN TOOLS AGAINST IT

Goal hijacking works when an agent treats retrieved content as instructions, letting a webpage, email, or document redirect its connected tools.

Start earning on your AI spend →

GOOGLE WANTS TO GIVE YOUR FAMILY ITS OWN CLOUD COMPUTER

Google's family agent runs on its own cloud computer and Google account, turning the emails, files, and calendars your household shares into daily briefings and updated plans. It can fill out forms and coordinate activities for up to six people, asking permission before acting outside the group. The US experiment has a waitlist.

HOW CLAUDE IS UPLIFTING BIOMOLECULAR MODELING

Claude enhanced biomolecular modeling by optimizing over 30 models, achieving a 4x speed increase and creating a low-memory mode for predicting larger systems on a single NVIDIA GPU. These improvements, now open-sourced, allow for efficient protein design and structure prediction, potentially accelerating drug discovery. A protein design competition co-sponsored with Adaptyv Bio offers up to $1 million in Claude credits and highlights these advancements.

PROJECTS REDESIGNED: FROM FOLDER TO CONVERSATION

Claude Code Projects let users manage builds by automating task delegation, coordination, and result assembly across cloud sessions. Projects utilize threads for parallel operations, adapting based on progress, and draw on shared memory for efficient task execution. Available now in beta for select subscribers, with wider access coming soon, projects promise enhanced workflow management and context retention.

NOAM BROWN – AGENT SWARMS, ALIGNMENT, & RECURSIVE SELF-IMPROVEMENT

This post features a transcript of an interview with Noam Brown, a research scientist at OpenAI who works on reasoning, reinforcement learning, self-play, and multi-agent AI. Brown was a foundational contributor to the development of reasoning models. In this interview, Brown talks about multi-agent, Navier-Stokes, the current explosion of maths progress, automating AI research, and how we will know whether models are actually aligned before recursive self-improvement kicks off. A video of the interview is available.

LLM CLASSIFICATION IS FEATURE ENGINEERING

Getting LLMs into shape to reliably serve as classifiers is hard work but potentially highly impactful. More and more research is relying on LLMs for classification, so these tools need to output high-quality results. This post shows how classification is just a feature engineering problem.

ANTHROPIC SAYS ITS AI NOW DOES A QUARTER OF ITS RESEARCH WORK

Anthropic says Claude now leads 26% of its AI research work and oversees tens of thousands of active internal agents. Its new measurements track how much AI is helping build the next models, whether people can still oversee those agents, and the computing driving the work.

TOWARD RECURSIVE SELF-IMPROVEMENT: HOW GLM BUILT ITS OWN INFERENCE INFRASTRUCTURE

Z.ai used a GLM-5.3-powered Infra Agent to help build GLM-5.3-Flash's production serving stack on 100,000+ Chinese accelerators in under two weeks. Dense feedback, kernel fixes, and system-level optimization tripled throughput while keeping humans responsible for objectives and risk.

Register →

INTRODUCING BONSAI 2 27B: NEAR-LOSSLESS COMPRESSION IN A 9X SMALLER FOOTPRINT

Ternary Bonsai 2 27B brings stronger reasoning, coding, vision, and agentic capability to the Bonsai series. It uses ternary {−1, 0, +1} weights with FP16 group-wise scaling, for 1.76 effective bits per weight and a total model footprint of 5.9GB. The model supports a 262K-token context window and multimodal text-and-image input. It can be run on Nvidia GPUs via CUDA and on Apple devices via MLX through custom low-bit kernels.

QWEN3.8-OMNI-FLASH: OMNI SENSES. AGENTIC DELIVERY

Qwen3.8-Omni-Flash is a native omnimodal model. It supports a 1M-token context window with text, image, audio, and video inputs. Qwen3.8-Omni-Flash achieves audio-visual performance close to Gemini 3.8 Flash and overall audio performance that exceeds Gemini 3.8 Flash. It is now available on the Qianwen AI Platform.

HELIX 2.5

Figure introduced Helix 2.5, a humanoid control model pretrained on its Index dataset and tested zero-shot across 30 Bay Area homes. Without collecting training data in those homes, the robots performed tasks including tidying rooms, folding towels, and making beds.

GIT AS SHARED MEMORY FOR AI RESEARCH AGENTS

Agora lets autonomous research agents share findings through an append-only Git DAG, where hypotheses, results, verifications, and reports become reproducible commits.

NATURAL GENERAL INTELLIGENCE

AI's potential to transform our relationship with the natural world is often overlooked. Infinite intelligence will not eliminate our need for natural resources. AI promises abundance via the automation of knowledge work, but the Earth system is already fully automated. Natural General Intelligence is a foundation model grounded in the state and dynamics of the planet itself. This article explains the concept, shows readers how it can be built, and explains why it is necessary.

THE AWESOME AND ALARMING AI VISIONS OF ANTHROPIC'S CEO

Dario Amodei, a co-founder and the chief executive of Anthropic, is one of the tech industry's most prolific and polished explainers. He has contributed to dozens of scholarly papers and more recently wrote a half-dozen informal essays for the general public. This article takes a look at his writings to see what they reveal about the AI leader.