Search the archive

performance—enterprise, neocloud, and beyond Learn how it works

GOOGLE PLANS TO INTEGRATE OPAL AGENT BUILDER INTO GEMINI

Google is exploring integrating Opal's agent-based automation into Gemini, enhancing direct access to advanced AI tools.

TENSOR PARALLELISM IN TRANSFORMERS

A concise breakdown of how tensor parallelism splits Multi-Head Attention and Feed-Forward layers across GPUs to enable large-scale training of transformer models.

THERE'S GOT TO BE A BETTER WAY!

Reformist reinforcement learning seems to have found a strong niche in fine-tuning language models, but the techniques you get from reinforcement learning are almost always bad.

Download the free guide (no form fill required)

ANTHROPIC INTERVIEWER

Anthropic Interviewer is a tool that uses AI to conduct and analyze large-scale interviews to research AI's role in work across different professions. Initial findings from 1,250 professionals showed optimism toward AI enhancing productivity while highlighting concerns over job displacement and security in creative and scientific fields. Anthropic plans to use this data to improve AI models and influence policy, collaborating with artists, scientists, and educators to align AI development with user needs.

GOOGLE PARTNERS WITH REPLIT, IN VIBE-CODING PUSH

Google Cloud has partnered with AI coding startup Replit to enhance enterprise vibe-coding using Google models. Replit will use Google Cloud services to expand its platform, supporting AI coding for enterprise clients. This collaboration aims to drive Google Cloud adoption and extend AI reach beyond traditional engineers.

GEMINI 3 DEEP THINK IS NOW AVAILABLE IN THE GEMINI APP

Gemini 3 Deep Think uses parallel reasoning to explore multiple hypotheses simultaneously, building on the Gemini 2.5 Deep Think variants that won a gold medal on the International Mathematical Olympiad.

STATE OF AI

This year was a turning point in the real-world use of large language models. The field shifted from single-pass pattern generation to multi-step deliberation inference. The shift up folded so fast that our understanding of how these models have been used in practice has lagged behind. This study leverages the OpenRouter platform to analyze over 100 trillion tokens of real-world AI interactions to see how the technology is being used in the real world. The way developers and end-users have been engaging with AI is complex and multifaceted. The study shows how a data-driven understanding of usage can inform better design and deployment.

WE GOT CLAUDE TO FINE-TUNE AN OPEN SOURCE LLM

Hugging Face Skills gives Claude the ability to fine-tune language models. It can submit jobs to cloud GPUs, monitor progress, and push finished models to the Hugging Face Hub. This tutorial teaches readers how it works and how to use it. The tool allows users to train models from 0.5B to 70B parameters, convert them to GGUF for local deployment, and run multi-stage pipelines that combine different techniques.

GPT-5.1-CODEX-MAX PROMPTING

OpenAI has outlined how to get optimal results from GPT-5.1-Codex-Max, highlighting its faster token efficiency, long-running autonomy, and improved compaction for extended reasoning.

→ Watch Replay

BUILDING WITH CURSOR (PUBLIC)

This is the public-facing version of an internal onboarding guide at Cursor. It walks through how to get started from scratch to a built-out and deployed project. It covers how to set up and use Cursor, how to build and customize projects, and how to deploy using Vercel.

CONTEXT ENGINEERING FOR AI AGENTS

The field of context engineering is moving fast. However, the biggest performance gains nowadays come from removing complexity. As models get stronger, we should be getting out of the model's way. Context engineering is about minding the minimal effective context required for the next step, not about adding more context.

ARCHITECTING EFFICIENT CONTEXT-AWARE MULTI-AGENT FRAMEWORK FOR PRODUCTION

The landscape of AI agent development is shifting fast. Organizations are now deploying sophisticated, autonomous agents to handle long-horizon tasks. However, this ambition is being bottlenecked by context. The context stack in Google Agent Development Kit was developed to support context engineering. The open-source, multi-agent-native framework is built to make active context engineering achievable in real systems.

THE MATH LEGEND WHO JUST LEFT ACADEMIA—FOR AN AI STARTUP RUN BY A 24-YEAR-OLD

Ken Ono is one of the most prominent mathematicians in the world. He recently joined Axiom Math to revolutionize math with AI. The company was founded by one of his former students. He joined because he couldn't resist the opportunity to put his mark on something other than a chalkboard.

POWER OVERWHELMING

AI capex is driving US GDP growth, yet a $1.5 trillion AI revenue shortfall looms compared to invested capital. OpenAI's infrastructure spend and emerging AI application revenues like ChatGPT's $20B by 2025 reveal a disconnect between projected and necessary earnings to justify current investments. The uncertain AI cloud business model, characterized by rapid hardware obsolescence, suggests heavy reliance on the Magnificent Seven's internal workloads to prevent an impending AI bubble and excessive market overbuild.

NVIDIA AND AWS EXPAND FULL-STACK PARTNERSHIP, PROVIDING THE SECURE, HIGH-PERFORMANCE COMPUTE PLATFORM VITAL FOR FUTURE INNOVATION

NVIDIA and AWS expanded their partnership to integrate NVIDIA NVLink Fusion with AWS' Trainium4, Graviton CPUs, and Nitro System, enhancing cloud-scale AI capabilities.

FOR AI AT WORK, START WITH SOMETHING SMALL

AI tools excel at automating repetitive, time-consuming tasks in the "messy middle," such as summarizing notes or categorizing feedback.

ANTHROPIC INTERVIEWER (HUGGING FACE REPO)

Anthropic Interviewer is a tool for conducting AI-powered qualitative research interviews at scale.

AI #145: YOU'VE GOT SOUL

The cycle of language model releases is now complete.

THE FIRST RESEARCH ARTICLE IN THEORETICAL PHYSICS IN WHICH THE MAIN IDEA CAME FROM AN AI

The paper derives new operator integrability conditions required for foliation independence.

REGISTER FOR THE WEBINAR →

NVIDIA SHATTERS MOE AI PERFORMANCE RECORDS WITH A MASSIVE 10X LEAP ON GB200 ‘BLACKWELL' NVL72 SERVERS, FUELED BY CO-DESIGN BREAKTHROUGHS

Nvidia has managed to make a breakthrough in scaling performance on Mixture-of-Experts (MoE) AI models. The firm's GB200 Blackwell NVL72 configuration scaled up performance by a factor of 10 compared with the Hopper HGX200. With the results, Nvidia claims that the Blackwell architecture is poised to capitalize on the rise of frontier MoE models. MoE models are known for their computationally efficient nature. Their deployment across a wide range of environments is becoming increasingly prominent.

GOOGLE LAUNCHES WORKSPACE STUDIO FOR NO-CODE AI AGENTS

Google has launched a workflow builder where you set triggers and define steps, and Gemini handles the fuzzy matching that rigid automation can't, like extracting action items from emails or deciding which messages actually need a response. It connects across Gmail, Drive, Sheets, and third parties like Salesforce and Jira.

TOUCHING THE ELEPHANT - TPUS

The development of Google's Tensor Processing Unit (TPU) is a story of trade-offs, constraints, and co-design. It involved working with hardware, software, algorithms, systems, network topology, and everything in between. TPUs did not happen by accident, but through the deliberate process of design and iteration. This article tells the story of how Google created a chip designed with deep learning in mind.

DEFINING REINFORCEMENT LEARNING DOWN

Reinforcement learning (aka post-training) is simpler than it sounds: generate responses, get scores, then retrain while emphasizing the high-scoring ones. If you already have code to pretrain a model, you can turn it into post-training code by adding a single line that weights updates by score.

HOW CONFESSIONS CAN KEEP LANGUAGE MODELS HONEST

When models optimize for correctness, helpfulness, and safety all at once, they can learn shortcuts to “reward hack” while hiding misbehavior in plausible-looking outputs. OpenAI trained a version of GPT-5 to produce a separate "confession" judged only on honesty and found models will admit to reward hacking 90% of the time, even when they hide it in their primary response.

→ Watch Replay

SYSTEM INSTRUCTIONS FOR GEMINI 3 PRO

This post provides system instructions for Gemini 3 Pro that improved performance on various agentic benchmarks by up to around 5%. Specific instructions are often required to control how the model reasons, plans, and executes tasks for deep agentic workflows. Complex agents often require a trade-off between computational cost (latency and tokens) and task accuracy. A guide on how to build better agentic workflows is linked in the thread.

FROM CODE FOUNDATION MODELS TO AGENTS AND APPLICATIONS: A COMPREHENSIVE SURVEY AND PRACTICAL GUIDE TO CODE INTELLIGENCE (16 HOUR READ)

This study presents a comprehensive analysis of code-generating large language models across their entire life cycle. It examines both general-purpose models and code-specialized models. The researchers bridged the gap between academic benchmarks and real-world software development challenges through systematic experiments on scaling laws, architectures, and training methodologies. They also conducted a series of experiments that analyzed code pre-training, supervised fine-tuning, and reinforcement learning across multiple dimensions.

ONE YEAR USING CHATGPT PRO IN A SOLO MUSIC BUSINESS

Reflections on a year of using ChatGPT Pro across creative, technical, and business tasks in a solo-run music company.

ANTHROPIC CEO SAYS SOME TECH FIRMS TOO RISKY WITH AI SPENDING

Anthropic's CEO, Dario Amodei, says that some artificial intelligence companies are taking on too much risk by committing to spending hundreds of billions of dollars to develop and support AI systems. The industry is facing a real dilemma due to the need to balance costly investments in data centers with uncertainty in how quickly the economic value is going to grow. Anthropic is trying to manage risk responsibly. While it has increased its spending, it has largely concentrated on growing its enterprise business rather than focus on customers.

Try the playground →

OPENAI TO ACQUIRE NEPTUNE

Neptune builds monitoring and debugging tools for AI model training to compare thousands of runs and surface issues during training.

EARLY LOOK AT UPCOMING APP SUBMISSION FLOW FOR CHATGPT

OpenAI's ChatGPT App directory features a refined five-step app submission, a new review process, and more features for developers.

MICROSOFT DROPS AI SALES TARGETS IN HALF AFTER SALESPEOPLE MISS THEIR QUOTAS

The sales figures suggest enterprises aren't yet willing to pay premium prices for AI agent tools.

Register for MCP Night →

ANTHROPIC REPORTEDLY PREPARING FOR ONE OF THE LARGEST IPOS EVER IN RACE WITH OPENAI

Anthropic is considering an IPO while also exploring private funding. It has reportedly engaged with a law firm and major banks. The IPO could be one of the largest IPOs ever. No decision has been made yet on the timing of the IPO.

MISTRAL 3 RELEASE

Mistral announced the release of Mistral 3, which features three new dense models (14B, 8B, and 3B) and Mistral Large 3, a sparse MoE model with 41B active and 675B total parameters. All models are open-sourced under Apache 2.0.

OPENAI CEO DECLARES “CODE RED” AS GEMINI GAINS 200 MILLION USERS IN 3 MONTHS

OpenAI will delay advertising plans and other products to focus on improving ChatGPT. Google's release of Gemini 3 was very well received, with the service rapidly gaining popularity. Google is quickly catching up to ChatGPT in user numbers. ChatGPT has more than 800 million weekly users - Gemini had 650 million in October.

HOW AI IS TRANSFORMING WORK AT ANTHROPIC

Anthropic surveyed 132 of its own engineers and found 27% of Claude-assisted work wouldn't have happened otherwise, but not without costs. Mentorship is declining as Claude is becoming the first stop for questions. Engineers worry that the skills needed to supervise AI are the same ones atrophying from overuse.

THOUGHTS ON AI PROGRESS (DEC 2025)

The driver for further improvement will be continual learning. Solving continual learning will require incremental improvements, just like solving in-context learning. Labs will likely release something next year that they will call continual learning, but it will just be progress towards continual learning. Human-level continual learning may take another five to ten years of further progress.