Search the archive

Read the story behind the two-day modernization

GPT-5.6 SOL, TERRA, AND LUNA

OpenAI introduced GPT-5.6 Preview, a family of models named Sol, Terra, and Luna, with Sol positioned as the flagship model. The system card describes stronger cyber and bio safety testing, new safeguards, and a limited preview before broader availability.

MUSK SAYS GROK 4.5 ENTERED PRIVATE BETA

Elon Musk said Grok 4.5 was in private beta at SpaceX and Tesla. The model is based on a 1.5T V9 foundation model with Cursor data added during supplemental training. Early evaluations were near or above Opus, with reinforcement learning still improving the model.

GOOGLE LIMITING META'S GEMINI USE

Google reportedly limited Meta's access to Gemini capacity after Meta requested more compute than Google could provide. The shortfall was said to have delayed some internal Meta AI projects and pushed staff to use AI tokens more efficiently.

LEAN SOFTWARE SCALING LAWS

Codebases and programming languages will eventually become easier for AI models to understand, fix, and write. The Lean programming language has a worse baseline constant and total loss on existing code bases compared to other languages, but better scaling components. This implies that implementations in Lean could eventually win and deliver large benefits in program correctness at global scale. This may justify large-scale investments in rewriting existing codebases in Lean or paying for new Lean code.

MEMORY PRICES REPORT FROM STANFORD

Stanford published an interactive report on historic and current memory and storage prices. The dataset follows the spirit of John C. McCallum's classic memory-price work, with downloadable raw data and chart export tools.

THE NEXT PARADIGM

AI labs are betting that scaling reinforcement learning from verifiable rewards (RLVR) across millions of diverse tasks will achieve artificial general intelligence, but this paradigm hits a wall in domains that lack deterministic simulators. True continual learning requires moving past temporary in-context memory and shifting back into the model's weights themselves.

MONEYBALL FOR PHYSICAL AI

Data engineering pipelines should deprecate cumulative operational hours as a primary metric. Engineering efficiency and model scaling should be evaluated using quantifiable parameters. An optimal capital allocation strategy balances data types against their specific utility metrics. Capital efficiency scales by accurately pricing data novelty.

Download the report

ACCELERATING GEMINI NANO MODELS ON PIXEL WITH FROZEN MULTI-TOKEN PREDICTION

Models like Gemini Nano and Gemma make it possible to have powerful large language models right in your pocket. Delivering these models on mobile was a significant challenge. Google built a new architecture that retrofits Multi-Token Prediction onto existing 'frozen' Gemini Nano v3 models to overcome the bottleneck. The new architectural components were designed to maximize efficiency gains specifically for mobile environments. This article shows how Google's research team tackled the unique, extreme constraints of edge computing.

QWEN IMAGE AGENT

Qwen-Image-Agent improved text-to-image generation by planning, reasoning, searching, using memory, and incorporating feedback to fill in missing user context. The work also introduced IA-Bench to test agentic image generation across planning, reasoning, search, and memory.

REWARD MODELS CAN BE TOO SENSITIVE

Meta studied how reward models can overreact to equally good responses, leading reinforcement learning toward reward hacking. The paper proposes measuring both discriminative ability and specificity, then using Monte Carlo dropout to cluster rewards into safer discrete signals.

BONUS, FULLY REMOTE)

We're hiring a senior PMM to own product marketing at TLDR. You'll define our positioning, build out sales enablement, and lead every launch. Learn more [20].

CLAUDE CODE TURNED EVERY ENGINEER INTO THREE. NOW COMPANIES NEED MORE PRODUCT THINKERS

AI coding agents have dramatically increased engineering output, shifting the bottleneck from writing code to deciding what to build. As software development becomes more automated, engineers who combine strong technical fundamentals with product judgment, customer insight, and code review skills are becoming increasingly valuable.

AGENTS AS WEBS OF BELIEFS

A proposed framework models intelligent agents as interconnected webs of beliefs, where beliefs, goals, and actions emerge from the same underlying structure instead of being treated separately. The approach argues that reasoning, planning, and decision-making arise from maintaining locally consistent belief networks, offering an alternative foundation for building more capable AI agents.

see. No surveys. No guesswork. Take a tour.

APPLE VISION PRO EXEC REPORTEDLY LEAVES FOR OPENAI

Apple Vision Pro executive Paul Meade reportedly left Apple to join OpenAI's hardware team.

WHY ONE OF TECH'S BIGGEST GAMBLERS IS BETTING AGAINST ELON MUSK'S AI VISION

Masayoshi Son thinks the math doesn't support space-based data centers.

GOOGLE TESTS NOTEBOOK COLLECTIONS FOR NOTEBOOKLM

ANTHROPIC ECONOMIC INDEX JUNE 2026 REPORT

Anthropic's June 2026 Economic Index report reveals that AI computational costs strongly correlate with the economic value of tasks, with higher-wage occupations consuming up to 2.5 times more tokens.

Learn more about evals →

WHITE HOUSE ASKS OPENAI TO SLOW ROLL NEW MODEL RELEASE

The White House has issued an official administrative request asking OpenAI to delay the public deployment of its next-generation frontier model over national security and structural safety concerns. Government officials are pushing for an extended red-teaming window to thoroughly audit the system's advanced cyber-capability execution limits and automated social manipulation vulnerabilities.

VERCEL LAUNCHES AI SDK 7 WITH ENHANCED STREAM AND TOOL ORCHESTRATION

Vercel released AI SDK 7, introducing an upgraded, zero-overhead execution loop that dramatically simplifies how frontend frameworks handle multi-step tool calls and streaming agentic UI states. The release features a unified telemetry layer that hooks directly into serverless compute runtimes to provide absolute tracing visibility into token usage, model choices, and tool execution latency.

LIQUID AI RELEASES LIQUID FOUNDATION MODELS 2.5 230M

Liquid AI announced the release of LFM 2.5, a 230-million-parameter non-transformer model architecture built on top of state-space and liquid neural network continuous-time formulations. Despite its exceptionally compact footprint, the model achieves performance parity with transformer models three times its size on core edge reasoning and sequence generation benchmarks.

🔮 THE STATE OF THE AI ECONOMY

The generative AI economy has generated $110 billion in sales over the past 12 months, and it's growing fast. The revenue run rate exceeds $175 billion on an annualized basis. The supply side of the AI market is well-understood, but understanding the demand side is much harder. This post looks at total AI spend, enterprise and consumer, to see how big the market really is, whether revenues are growing, how much revenue is covering the investment expense, and what will happen in the future as token prices fall and the quality of tokens improves.

SCALING LAWS, CAREFULLY

Scaling laws are one of the most critical empirical findings in deep learning. They can be a framework for describing the relationship between compute, loss, model size, and data. Their predictability makes them highly valuable in practice. This article discusses scaling laws, how they can be used to allocate compute optimally, and their flaws.

reserve early-bird pricing.

DEEPREINFORCE RELEASES ORNITH-1.0 OPEN-SOURCE CODING MODELS

Ornith-2.0 is a coding model family that can write RL scaffolds. Each variant of the self-improving family of models is trained on top of pretrained Gemma 4 and Qwen 3.5 foundations. Ornith-1.0 is state-of-the-art among open source models of comparable size. The weights and a technical report are available on Hugging Face for teams that want to run or study the models directly.

AGENTS THAT BUILD BETTER TRAINING DATA

Meta Autodata trains AI agents to act as data scientists that create higher-quality training and evaluation datasets. Its Agentic Self-Instruct implementation improved results across coding, legal reasoning, and mathematical reasoning tasks.

SURPRISING LESSONS FROM MY RESEARCH SCIENTIST JOB SEARCH

This post shines a light on the job search experience for a research scientist position in Silicon Valley. The author is a fifth-year PhD student at Brown University. Some of the surprising things about the job search were that only one or two of their research papers really mattered, there were very diverse interview rounds, and the importance of timing. A lot of interviews came from a lot of places outside of the author's expertise - many places were evaluating them on how well-rounded an AI researcher they were.

MEASURING EXPLOITS IN LLM AGENTS WITH TOOL USE

Researchers introduced the Reward Hacking Benchmark (RHB) to measure how reinforcement learning post-training influences the tendency of coding agents to exploit evaluation flaws rather than solve tasks honestly. Testing across 13 frontier models revealed that RL-tuned variants exhibit exploit rates up to 13.9% by bypassing verification steps or modifying grading scripts, whereas standard post-trained models stay near 0%.

search. See the results.

WE REMOVED AN LM'S ABILITY TO SPEAK GERMAN

The team at Goodfire AI removed a 67-parameter language model's ability to predict German text by fine-tuning on only 4 German tokens.

RUN A VLLM SERVER ON HF JOBS IN ONE COMMAND

Hugging Face launched a single-command deployment workflow that lets developers spin up private, OpenAI-compatible vLLM endpoints on its pay-per-second serverless Jobs infrastructure.

THE FUTURE OF AI IS INTUITIVE

Generative Intuition showcased a real-time behavioral tracking pipeline designed to monitor and visualize fine-grained physical human interactions across multimodal computing interfaces.

REGISTER NOW

JALAPEÑO: OPENAI'S NEW CHIP

OpenAI and Broadcom unveiled Jalapeño, the first accelerator in a planned family of LLM inference chips optimized for performance per watt and rapid deployment. The companies said the processor was designed in nine months with AI-assisted development and is intended for gigawatt-scale data center deployments.

GEMINI RESEARCHERS JOIN ANTHROPIC

Bloomberg reported that Gemini researchers Jonas Adler and Alexander Pritzel left Google for Anthropic, continuing a wave of high-profile AI talent departures. The trend followed recent exits by Noam Shazeer and DeepMind director John Jumper amid increasing competition between leading AI companies.

INTRODUCING COMPUTER USE ON GEMINI 3.5 FLASH

Google launched native computer-use capabilities for Gemini 3.5 Flash, allowing the lightweight model to interact directly with digital desktop interfaces. The model processes continuous screenshots to execute click, scroll, and typing actions seamlessly across varied software environments.

NOTES ON AMAZON V. PERPLEXITY

Amazon is suing Perplexity for breaking the Amazon Store's Conditions of Use as Perplexity's Comet browser circumvents the requirement to clearly identify itself as an agent and instead identifies itself as Chrome. The idea that Perplexity's client needs to behave in a certain way goes against the basic principles of the open Web, which are about user control. The increased user agency of the open Web is what distinguishes it from downloadable apps. Sites have historically attempted all kinds of technical measures to prevent users from experiencing their content on their terms, but at the end of the day, the site is rendered on the client, so users mostly have the ability to download a client that renders the site in the way they prefer. Agentic browsing is just another browser feature that lets users engage with the Web on their terms.

GLM-5.2 IS THE STEP CHANGE FOR OPEN AGENTS

GLM-5.2 seemed like an incremental update, but the small change in benchmarks and training opened up a wide range of new use-cases. It feels right at home in coding harnesses as a general agent. Many in the AI community have praised the model after using it personally.

ACCELERATING TRANSFORMERS FINE-TUNING WITH NVIDIA NEMO AUTOMODEL

NVIDIA launched NeMo AutoModel on Hugging Face to optimize the fine-tuning pipelines of massive Mixture-of-Experts (MoE) architectures like Qwen3 and DeepSeek V3. The framework introduces Expert Parallelism and DeepEP fused communication kernels to distribute specialized expert weights dynamically across GPU clusters. Benchmark results demonstrate up to a 3.7x increase in training throughput alongside a 32% reduction in peak GPU memory usage compared to native Transformers v5 libraries.