Search the archive

GROK 4 RELEASE LIVESTREAM

xAI will hold a livestream for the Grok 4 release on Wednesday at 8 PM PT.

APPLE'S TOP AI EXECUTIVE RUOMING PANG LEAVES FOR META

Ruoming Pang, Apple's head of AI models, is leaving to join Meta's new superintelligence team. Meta is aggressively recruiting top AI talent with multi-million-dollar packages and recently reorganized its AI efforts under Meta Superintelligence Labs, led by Alexandr Wang.

CHATGPT EXPERIMENTS WITH ‘STUDY TOGETHER' MODE

Some users have reported a new experimental ChatGPT feature called "Study Together" that encourages more interactive learning by prompting users with questions instead of direct answers. Details remain sparse, and OpenAI has not made an official announcement.

ECONOMICS OF CLAUDE 3 OPUS INFERENCE

Anthropic recently announced the deprecation of API access to Claude 3 Opus. The decision to deprecate the API addresses a real operational issue. This article looks at the economics of reduced-scale inference and looks at alternative solutions that can be a win-win for both Anthropic and independent researchers. Keeping inference access to Claude 3 Opus is more complex than it seems at a glance.

A REVIEW OF ALPHA SCHOOL, THE PRIVATE SCHOOL WITH 2-HOUR DAYS AND AI TEACHERS

A parent's year-long investigation revealed the $40,000 Austin school actually takes 3.5 hours daily with a 5:1 teacher ratio and extensive incentive systems, not AI-powered teacher-free education as marketed. While students do advance 2.6x faster through material using personalized learning platforms, parents argue the real value isn't speed, but freeing up time - it potentially gives kids ~9 additional years outside of the classroom to pursue their own interests.

your users find it. See, try, or read

MICROJAX

Microjax is a tiny autograd engine with a Jax-like API. It was inspired by Andrej Karpathy's Micrograd, a PyTorch-like library with about 150 lines of code. JAX uses a more functional style, which some developers prefer.

CORT (CHAIN OF RECURSIVE THOUGHTS)

CoRT enhances AI models by making them self-evaluate and repeatedly generate alternatives for optimal responses. Testing with Mistral 3.1 24B showed significant performance improvements in programming tasks. The process involves iterative response generation and selection, leading to more refined AI outputs.

BITNET

An inference framework for Microsoft's BitNet b1.58, a 1.58-bit (ternary) large language model designed for efficient and lossless CPU inference using optimized low-bit kernels.

THE ‘CHATGPT MOMENT' IN ROBOTICS AND BEYOND

Three years ago, getting a robot to reliably pick up objects required an army of engineers. Today, a college student can download an open-source vision-language-action model, fine-tune it on a weekend, and achieve results that would have taken industry teams months to accomplish. This article looks at what a 'ChatGPT moment' for robotics would look like, the current state of the industry, upcoming technologies, and likely winners. Being surrounded by robots in everyday settings will feel surreal at first, but we will soon rely on them just like we do with AI assistants for the most basic needs of our lives.

REPLIT DYNAMIC INTELLIGENCE FOR REPLIT AGENT

Replit introduced Dynamic Intelligence for its Agent, adding three capabilities: Extended Thinking, High Power Model, and Web Search. These enhancements improve context awareness, iterative reasoning, and autonomous behavior, allowing the Agent to adapt and solve complex tasks efficiently. Users can toggle these features per request, optimizing the Agent's problem-solving capabilities for different scenarios.

HUAWEI DENIES ALLEGATIONS OF COPYING ALIBABA'S AI MODEL

An anonymous whistleblower claiming to work at Huawei presented a technical analysis showing that the new Pangu model was trained by “upcycling” Alibaba's Qwen model rather than training from scratch. Using "model fingerprinting", they calculated a 0.9 correlation coefficient between the models. The evidence isn't conclusive - critics have pointed out flaws in the fingerprinting methodology.

COREWEAVE TO ACQUIRE CORE SCIENTIFIC IN $9 BILLION ALL-STOCK DEAL

The AI cloud infrastructure provider will buy the data center operator to eliminate $10 billion in future lease obligations and gain ownership of 1.3 gigawatts of compute capacity.

MIRAGE, AN AI-NATIVE GAME ENGINE FOR REAL-TIME WORLD GENERATION

The first generative game engine lets players modify environments through natural language during gameplay, launching with GTA-style and racing demos that run entirely on AI-generated content at 16 FPS.

CURSOR APOLOGIZES FOR UNCLEAR PRICING CHANGES

Cursor's parent company, Anysphere, issued an apology after rolling out pricing changes to its Pro plan without sufficient clarity, prompting backlash from its user base.

GOOGLE EXPLORES AI IN MENTAL HEALTH TREATMENT

Google announced two mental health AI initiatives: a practical guide to responsibly deploying AI in mental health care, and a multi-year research partnership with DeepMind and Wellcome Trust to study AI-driven diagnosis and treatment of anxiety, depression, and psychosis.

GROK 4 BENCHMARKS LEAK WITH 45% SCORE ON HUMANITY LAST EXAM

Leaked benchmarks show that Grok 4 will be state-of-the-art. References to the model have surfaced in the xAI console. If the benchmarks are real, Grok 4 could outperform leaders like Gemini 2.5 Pro, o3 Pro, and Claude 4 Opus. xAI faces competitive pressure to release Grok 4 before the market shifts again - OpenAI, Google, and Anthropic are rumored to be preparing fresh releases for launch.

CHARACTER AI'S REAL-TIME VIDEO GENERATION

Character.AI's TalkingMachines is a real-time, audio-driven video generation model that creates FaceTime-style animations from a single image and voice input.

ADDING MEMORY TO GEMINI 2.5 CHATBOTS

A guide explaining how to use the Gemini API and the open-source mem0 tool to give Gemini 2.5 chatbots long-term memory. The setup allows bots to recall past interactions, personalize responses, and reduce repetition for more context-aware conversations.

THE AMERICAN DEEPSEEK PROJECT

Meta's recent AI shortcomings have created a vacuum in the open-source AI ecosystem that has largely been filled by Chinese models. If current dynamics continue, the AI world will be split between powerful but expensive, closed-source American models and low-cost, ubiquitous but potentially compromised Chinese models. The US likely has a very small window, roughly the next two years, to counter this trend by investing $100-500 million into an open-source model that is as good as the best closed-source model.

BABA IS EVAL

'Baba is You' is a puzzle game where the rules have to be manipulated to win. The level of abstraction required to solve most levels makes it a formidable reasoning benchmark. This study looks at how large language models fare in playing the game. At the moment, Claude 4 is pretty bad at playing the game - a reasoning model may be better equipped to play, so the next step in the study may be to test those models.

datacenter. Get the datasheet

WHAT CAN AGENTS ACTUALLY DO?

There's a lot of excitement about what AI can do, but many of the conversations about the technology are so abstract that they border on meaningless. This post attempts to concisely summarize how AI agents work. It looks at a handful of real-world use cases for the technology. AI agents are a multiplier on software quality and system design. If software or systems are poorly designed, agents will only cause harm.

SAKANA AI'S TREEQUEST: DEPLOY MULTI-MODEL TEAMS THAT OUTPERFORM INDIVIDUAL LLMS BY 30%

Sakana AI's TreeQuest uses Multi-LLM AB-MCTS to combine multiple LLMs. Achieving a 30% performance boost over individual models, this technique leverages the unique strengths of different models through Adaptive Branching Monte Carlo Tree Search, dynamically assigning the optimal model for each task. Available as an open-source tool, TreeQuest allows businesses to apply this approach to complex problems, enhancing AI capabilities and reducing hallucination risks.

AGENT-SQUAD

A framework for building collaborative multi-agent AI systems that can plan, delegate, and work together to solve complex tasks.

WHY I DON'T THINK AGI IS RIGHT AROUND THE CORNER

It's hard to get normal humanlike labor out of LLMs because they lack some fundamental capabilities. They don't get better over time, and this lack of continual learning is a huge problem. There's no way to give models human-level feedback, so users are stuck with the abilities they have out of the box. Messing around with system prompts doesn't produce anything close to the kind of learning and improvement that human employees experience. Humans are able to build up context, interrogate their own failures, and pick up small improvements and efficiencies as they practice a task.

GOOGLE FACES EU COMPLAINT OVER AI OVERVIEWS

The Independent Publishers Alliance filed an antitrust complaint with the European Commission, alleging that Google's AI Overviews misuse web content and have led to traffic and revenue loss for publishers.

Subscribe now.

ELON MUSK CONFIRMS XAI IS BUYING AN OVERSEAS POWER PLANT TO POWER ITS NEW DATA CENTER

xAI's next data centers are expected to house millions of AI chips.

HOLY SMOKES! A NEW, 200% FASTER DEEPSEEK R1-0528 VARIANT APPEARS FROM GERMAN LAB TNG TECHNOLOGY CONSULTING GMBH

24-year-old German firm TNG Technology Consulting GmbH's DeepSeek-TNG R1T2 Chimera delivers a notable boost in efficiency and speed while using significantly fewer output tokens.

NFDG: THE $1.1B VC FUND THAT 4X'D IN TWO YEARS—THEN GOT ACQUIRED BY META

This post looks at NFDG's portfolio, advisory board, performance, success factors, and more.

NVIDIA'S DEAL TO BUY CANADIAN AI STARTUP CENTML COULD TOP US$400M

CentML makes software that operates between users' AI models and the chips powering them, making the systems run better.

GROK 4 SPOTTED AHEAD OF LAUNCH WITH SPECIAL CODING FEATURES

Grok 4 (grok-4-0629) offers unparalleled performance in natural language, math, and reasoning.

RESEARCHERS SEEK TO INFLUENCE PEER REVIEW WITH HIDDEN AI PROMPTS

Researchers are embedding hidden AI prompts in academic papers on arXiv to influence peer reviews positively.

ILYA SUTSKEVER BECOMES CEO OF SAFE SUPERINTELLIGENCE AFTER META POACHED DANIEL GROSS

Ilya Sutskever, OpenAI co-founder, has become the CEO of Safe Superintelligence after Meta hired former CEO Daniel Gross. Meta, led by CEO Mark Zuckerberg, will continue its AI expansion with significant investments and the creation of Meta Superintelligence Labs. Sutskever will commit to maintaining Safe Superintelligence as an independent entity, dismissing acquisition attempts by Meta.

COREWEAVE IS THE FIRST CLOUD PROVIDER TO DEPLOY NVIDIA'S LATEST AI CHIPS

CoreWeave is the first cloud provider to deploy Nvidia's next-gen Blackwell Ultra AI chips, supplying them in Dell-built systems. The systems use 72 Blackwell Ultra GPUs and 36 Grace CPUs, all liquid-cooled and U.S.-assembled.

META IS REPORTEDLY TRAINING ITS AI CHATBOTS TO SEND UNPROMPTED MESSAGES

Meta's AI chatbots, developed with its no-code AI Studio, are being trained to send unprompted follow-up messages, enhancing user retention and engagement. The project, internally called "Project Omni," allows users to customize chatbot appearances and content across Meta platforms. While designed to boost engagement and revenue, these proactive messages will only occur if initiated by the user and adhere to established conversational guidelines.

CODING AGENTS 101: THE ART OF ACTUALLY GETTING THINGS DONE

The Devin creators advocate "defensive prompting" - explicitly telling agents HOW to approach tasks, not just what to build - while noting 1-6 hour tasks offer the highest ROI. They observe that senior-to-staff engineers adopt AI tools the fastest because they understand how to architect tasks for junior employees.

THE END OF MOORE'S LAW FOR AI? GEMINI FLASH OFFERS A WARNING

The AI industry has operated under its own version of Moore's Law over the past few years with an unwavering belief that the cost of intelligence would perpetually decrease by orders of magnitude each year. Each new model generation was not only more capable, but also cheaper to run. Google broke that trend last week with the introduction of Gemini 2.5 Flash. The input token price for the model doubled while the output price more than quadrupled. This is the first time a major provider has backtracked on the price of an established model, and it may signal a turning point - the industry may no longer be on an endless downward slide of cost.

CONTEXT ENGINEERING FOR AGENTS

LangChain shares a detailed guide on “Context Engineering,” a key part of agent building, including popular patterns and how to implement them

INFERENCE-TIME SCALING AND COLLECTIVE INTELLIGENCE FOR FRONTIER AI

The new inference-time scaling method combines o4-mini, Gemini-2.5-Pro, and DeepSeek-R1 to achieve 30% on ARC-AGI-2 benchmark, outperforming individual models by dynamically selecting which AI handles each problem. The approach extends Monte Carlo Tree Search to balance generating new solutions versus refining existing ones, with the open-source TreeQuest framework enabling practical deployment.

OPEN SOURCE RL LIBRARIES FOR LLMS

Anyscale researchers compare TRL, Verl, OpenRLHF, and six other frameworks across adoption metrics, system properties, and technical architecture to help developers choose the right tool for RLHF, reasoning models, or agentic training scenarios.

APPLYING RL: IMPROVING CODE MERGING

Osmosis-Apply-1.7B, a fine-tuned version of Qwen3-1.7B using reinforcement learning, outperforms larger foundation models like OpenAI o3 in code merging tasks with a reward score of 0.9893 and at a significantly lower cost. Trained on a subset of the CommitPackFT dataset, its efficiency stems from using GRPO with FSDP strategy, optimizing for successful code merges without KL divergence or entropy bonuses.

RUN AND FINETUNE GEMMA 3N

A guide on how to run Google's new Gemma 3n locally with Dynamic GGUFs on llama.cpp, Ollama, or Open WebUI, and how to fine-tune with Unsloth.

CHATGPT REFERRALS TO NEWS SITES ARE GROWING, BUT NOT ENOUGH TO OFFSET SEARCH DECLINES

Users are increasingly getting their news directly from AI or AI-powered search results. The number of news searches on the web that resulted in no click-throughs since the launch of Google's AI Overviews has grown from 56% to nearly 69% as of May this year. News-related prompts in ChatGPT grew by 212% from January 2024 through May 2025. Visibility in Google Search results and good SEO practices may no longer deliver the value they did before AI.

Subscribe now.