Search the archive

GEMMA MULTIMODAL FINE-TUNER

Gemma Multimodal Fine-Tuner lets Mac users fine-tune Gemma on text, images, and audio on remote data.

WHEN WILL ANTHROPIC SURPASS NVIDIA?

Anthropic reached $10 billion in revenue in under four years, far outpacing other software companies like ServiceNow and Shopify.

Add the latest episode to your Spotify / Apple / Youtube queue

Listen to learn

* Why product-market fit is incomplete without pricing-product-market fit * How usage-based models demand a completely different operating model across product, sales, and finance * The pricing innovations brands like HUGGINGFACE and NVIDIA are adopting

OPENAI TESTS NEXT-GEN IMAGE V2 MODEL ON CHATGPT AND LM ARENA

OpenAI is testing its next-gen Image V2 model on ChatGPT and LM Arena in three variants. Early tests show improvements in UI design rendering, prompt adherence, and compositional understanding. The outcome of these tests could bolster OpenAI's competitive position against Google's offerings.

GOOGLE TESTS JULES V2 AGENT CAPABLE OF TAKING BIGGER TASKS

Google is developing the Jules V2 ("Jitro") coding agent, designed to autonomously manage high-level development goals rather than specific tasks. Launching via waitlist, it aims to redefine AI software development by shifting focus from task-based commands to KPI-driven outcomes. The approach might benefit teams handling large codebases, but it faces challenges of unpredictable changes and trust issues.

ANTHROPIC EXPANDS PARTNERSHIP WITH GOOGLE AND BROADCOM FOR MULTIPLE GIGAWATTS OF NEXT-GENERATION COMPUTE

Anthropic has signed an agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity expected to come online starting in 2027. The capacity is necessary to serve the exponential growth in Anthropic's customer base. It will also enable Claude to define the frontier of AI development. The vast majority of the new compute will be sited in the US.

SOME OF META'S NEW AI MODELS WILL EVENTUALLY BE OPEN-SOURCE

Meta is close to releasing its new AI models. Some of the models will eventually be released with an open-source license. Meta plans to focus on consumers rather than the enterprise market. It previously embraced open-source AI with its Llama models, and CEO Mark Zuckerberg has written a manifesto declaring open-source AI as the past for ared. The company will be pursuing a hybrid strategy of proprietary and open-source models going forward.

AIS CAN NOW OFTEN DO MASSIVE EASY-TO-VERIFY SWE TASKS AND I'VE UPDATED TOWARDS SHORTER TIMELINES

There is now an almost 2x higher probability of full AI R&D automation by the end of 2028. Strong performance in easy-and-cheap-to-verify SWE tasks that don't require new ideas will allow AI to substantially speed up AI R&D. A lot of the work in AI R&D involves these types of tasks, and AI can complete them significantly cheaper and faster than humans.

ANATOMY OF MERCOR'S DATA BREACH

Mercor is an AI-powered talent marketplace that connects software engineers, AI data labelers, and knowledge workers with companies seeking contract labor. The company claims it was recently breached due to a supply-chain attack on a popular open source package. However, there are several reasons to doubt that claim. Saying that it was affected by the supply-chain attack may be an attempt to deflect from deeper, more embarrassing security failures.

OPENAI'S $122B “VC ROUND” IS VENDOR DEALS, CONTINGENT CAPITAL, AND A GUARANTEED RETURN IT ARGUABLY CAN'T AFFORD

OpenAI closed a $122 billion investment round at an $852 billion post-money valuation at the end of last month. Amazon, Nvidia, and SoftBank accounted for $110 billion of that investment, while the remaining $12 billion came from a broader pool of institutional and individual investors. This article breaks down the numbers of the deal. Completely separate from the round, OpenAI is negotiating a joint venture with several private equity firms valued at roughly $10 billion pre-money with a guaranteed minimum return of 17.5%, plus early access to OpenAI models before general availability.

THE ANATOMY OF AN AGENT HARNESS

This post breaks down the agent harness, the full software infrastructure that transforms an LLM into a goal-directed, tool-using agent. Production harnesses use layered memory, verification loops, subagent orchestration, and structured tool execution to prevent context loss, errors, and silent failures. Frameworks like Claude Code, OpenAI Agents SDK, and LangGraph show that harness design alone can shift agent performance by 20+ ranks, making it the critical engineering layer around any LLM.

TURNING THE WEB INTO A FILESYSTEM

Code hallucinations are mostly a data problem, caused by outdated or fragmented training data. Nia's approach mounts documentation sites as virtual filesystems, letting agents grep, cat, and tree through docs in real time, so they code against current APIs rather than stale embeddings. This client-side, in-memory shell works with Claude Code, Codex, Copilot, Gemini, and OpenCode, providing a universal, low-latency interface for agents to browse structured information without complex tool schemas or container overhead.

GITNEXUS

GitNexus indexes entire codebases into knowledge graphs, making AI agents like Cursor and Claude Code aware of every relationship, from dependencies to call chains. It offers a CLI for deep integration with editors and a web UI for visual exploration, both providing comprehensive context and reducing the likelihood of blind edits and breaking changes.

WHAT TO KNOW ABOUT OPENAI'S IDEAS FOR A WORLD WITH ‘SUPERINTELLIGENCE'

OpenAI recently released policy proposals for a world with AI that far surpasses human capabilities. The proposals suggest creating an AI-centric tax system, a public investment fund focused on AI that could regularly distribute returns to Americans, and more. OpenAI has embraced President Trump's AI strategy, including his approach of limited guardrails for the industry, but it still wants to cooperate with the rest of the world in identifying dangerous AI systems and containing them. It supports giving the government a bigger role in evaluating models.

AI IS BECOMING AN OPERATING SYSTEM LAYER

AI must become an operating system layer across entire companies, reshaping product development, service delivery, and internal operations. This approach involves integrating AI into core functions rather than treating it as a mere feature. Companies not adopting this model risk inefficiency and losing competitive edge, making AI integration crucial for investment appeal.

Some AI tools remember your conversations. Littlebird

understands what you've been working on. It observes your screen and meetings, building a private memory that grows with you. Try it free.

OPENAI SAFETY FELLOWSHIP PROGRAM

OpenAI announced a new fellowship running from September 2026 to February 2027 that will support external researchers working on AI safety, alignment, and related technical challenges with mentorship, compute, and stipends.

OFFLINE AI DICTATION APP FROM GOOGLE

Google released an iOS app using on-device ASR models for dictation that offers live transcription, filler-word removal, and optional cloud-based editing via Gemini models.

OPENAI ASKS CALIFORNIA, DELAWARE TO INVESTIGATE MUSK'S 'ANTI-COMPETITIVE BEHAVIOR' AHEAD OF APRIL TRIAL

OpenAI urges California and Delaware to investigate Elon Musk's alleged anti-competitive behavior ahead of an April trial.

THE TERM “AGI” IS ALMOST USELESS AT THIS POINT

The gaps between the different conceptions of the term 'AGI' are starting to damage our ability to think together about how to navigate the transition to a world with extremely advanced AI.

Download the playbook for free

ANTHROPIC ENDED SUBSCRIPTION-BASED OPENCLAW USAGE

ANTHROPIC BUYS BIOTECH STARTUP COEFFICIENT BIO IN $400M DEAL

Anthropic acquired biotech AI startup Coefficient Bio for $400 million in stock to bolster its healthcare and life sciences ventures. Coefficient Bio, founded by former Genentech researchers, used AI to expedite drug discovery and biological research. The 10-person team will integrate into Anthropic's health and life sciences division.

CONTINUAL LEARNING FOR AI AGENTS

Learning within AI agents can happen at the model, harness, or context layers. Understanding the difference can change how systems that improve over time are built. The model layer is the model weights themselves. The harness is the code, instructions, and tools that drive the agent. Context is the additional context that lives outside the harness for more configuration. Most people jump to the model when discussing continual learning, but in reality, an AI system can learn at all three of these levels.

A TAXONOMY OF RL ENVIRONMENTS FOR LLM AGENTS

RL environments are training grounds for agents. Task distribution determines what skills agents develop, and harnesses control how they interact. Verifiers define what 'good' means. The state and configuration determine how realistic the training is.

LLM WIKI

This 'idea file', designed to be copied and pasted into an LLM agent, contains a pattern for building knowledge bases using LLMs. It helps LLMs incrementally build and maintain a persistent wiki that can be extended as the model continues learning. In this framework, the human curates sources, directs the analysis, asks questions, and thinks about what it all means, while the model does the rest. The agent makes edits based on conversations, and users can browse the changes in real-time.

I STILL PREFER MCP OVER SKILLS

The industry is pushing hard for Skills as the new standard for giving LLMs capabilities, but the Model Context Protocol (MCP) is a far superior, more pragmatic architectural choice. Skills are great for pure knowledge and teaching agents how to use an existing tool. MCPs give agents actual access to services, making them the right tool for the job in many cases.

EMBARRASSINGLY SIMPLE SELF-DISTILLATION IMPROVES CODE GENERATION

Simple self-distillation (SSD) is a process where models are fine-tuned on samples of their output with standard supervised fine-tuning. It offers a complementary post-training direction for improving LLM code generation. SSD can improve code generation in models using only raw outputs. This study looks at why such a simple method can work.

META-HARNESS: END-TO-END OPTIMIZATION OF MODEL HARNESSES

The performance of large language model systems depends not just on model weights, but also on their harnesses. Meta-Harness is an outer-loop system that searches over harness code for LLM applications. Its harnesses surpass the best hand-engineered harnesses on agentic coding benchmarks. Meta-Harness shows how richer access to prior experience can enable automated harness engineering.

THEREFORE I AM. I THINK

LLMs often decide actions before generating reasoning tokens, influencing their chain of thought. A linear probe can decode these decisions from pre-generation activations with high accuracy.

INTERACTION-AWARE VIDEO OBJECT REMOVAL

Netflix's VOID is a video inpainting model that removes objects along with their physical and visual interactions, using interaction-aware mask conditioning built on CogVideoX.

BINDING ACTIONS TO MULTIPLE SUBJECTS IN VIDEO

ActionParty introduces per-subject state tokens with spatial biasing to ensure correct action assignment across multiple entities in generated videos, addressing common failures in subject-action binding.

HOW TO BUILD YOUR SECOND BRAIN

A simple three-folder system (raw, wiki, and outputs) turns scattered notes into a structured, AI-maintained knowledge base using plain text files and a lightweight schema. Tools like agent-browser automate content ingestion, while AI compiles, links, and updates a personal wiki from raw inputs without manual organization. The system improves over time by saving outputs back into the loop and running periodic health checks to catch errors and gaps before they compound.

APPLE AT 50: THE IPHONE MAKER 'BLEW A 5-YEAR LEAD' ON AI, BUT FORMER INSIDERS SAY IT CAN STILL WIN

Apple's reliance on Google's Gemini AI for a revamped Siri marks a strategic shift as the company addresses its lag in generative AI compared to peers. This partnership raises concerns about user data management while highlighting Apple's bet on AI capabilities integrated into devices. As AI technology shifts towards the device, Apple aims to leverage its strengths in design and privacy to regain competitiveness in the AI landscape.

MARC ANDREESSEN INTROSPECTS ON THE DEATH OF THE BROWSER, PI + OPENCLAW, AND WHY "THIS TIME IS DIFFERENT"

AI, a result of "80-year overnight success," is not another hype cycle but a culmination of decades of research leading to significant breakthroughs in reasoning, coding, and self-improving agents.

META PAUSES WORK WITH MERCOR AFTER DATA BREACH PUTS AI INDUSTRY SECRETS AT RISK

Meta paused its work with Mercor following a security breach that may have exposed proprietary AI training data.

MICROSOFT'S AI IN ITS OWN TERMS: "USE COPILOT AT YOUR OWN RISK"

Microsoft's Copilot terms label it for entertainment only, cautioning users against relying on it for serious decisions.

25 PERCENTAGE POINTS.

That's a gap that compounds fast, at 75% per-step accuracy across a 5-step workflow, fewer than 24% of processes complete correctly.

Read on for more

CURSOR 3

Cursor released a redesigned interface focused on agent-driven development, enabling multi-repo workflows, clearer abstraction, and coordination between local and cloud agents.

QWEN3.6-PLUS: TOWARDS REAL WORLD AGENTS

Qwen3.6-Plus perceives the world with greater accuracy and sharper multimodal reasoning than previous models. It offers a highly stable and reliable foundation for the developer ecosystem and delivers a truly transformative 'vibe coding' experience. The model marks a critical milestone in the journey toward native multimodal agents. The Qwen team plans to release open-source, smaller-scale variants of the model in the coming days.

GEMMA 4 OPEN MODELS

Google DeepMind introduced Gemma 4, a new generation of open models optimized for reasoning and agent workflows, offering high performance per parameter under an Apache 2.0 license.