Next Week in The Sequence:

• Our series about model distillation continues.

• We have a surprising mega interview.

• We dive into the Astra, Fable and Muse Spark releases to keep you up to date.

• We will discuss the possible “ChatGPT moments” for robotics.

Subscribe and don’t miss out:

📝 Editorial: Model Madness: The Frontier Has a Refresh Button

The AI industry has developed a peculiar new benchmark: can you finish reading a model’s system card before its replacement ships? This week, OpenAI, Anthropic, Meta, and Google turned the release calendar into a competitive sport. Somewhere, an engineer is still updating last week’s model comparison spreadsheet. Please give them space.

OpenAI’s GPT-6 Astra arrives with an expansive pitch spanning computer use, software engineering, and scientific work. OpenAI reports 98% on FrontierMath Tier 4 alongside improvements in computer interaction. The practical ambition is clear: models that navigate software, execute complicated workflows, and deliver usable work with less supervision. OpenAI is also updating the Codex harness to accelerate computer use, underscoring how much performance depends on the combination of model and surrounding infrastructure. The benchmark chart is becoming a job description—and the software around the model is becoming part of the résumé.

Anthropic’s Claude Fable 5.1 and Mythos 5.1 push coding, knowledge work, and scientific research forward. A revealing detail: they share the same underlying model, with different safeguards and access arrangements. Fable is generally available; Mythos is restricted to trusted access programs. Anthropic estimates that cheaper cache reads will reduce costs for typical workloads by around 25%, with larger savings possible for highly agentic work. That matters when an agent repeatedly revisits a substantial working context. Capability, deployment policy, and inference economics increasingly arrive in the same announcement.

Meta’s Muse Spark 1.3 focuses on the unglamorous mechanics that make agents useful: maintaining requirements across long tasks, handling conflicting information, revising plans, and asking for help. Meta says it trained across multiple agent harnesses to improve generalization between environments. It also emphasizes keeping track of different tasks within a single conversation, including when users interrupt or redirect ongoing work. Anyone who has watched an agent confidently abandon the original task halfway through a workflow will appreciate the ambition. Remembering what you were hired to do remains an underrated capability.

Google supplied the week’s best illustration of the tempo: Gemini 3.8 Flash is its third Flash release in six weeks. Google reports stronger coding and reasoning while retaining 3.7 Flash’s speed and introductory pricing. Alongside it, Flash Cyber targets vulnerability discovery and patching through restricted access. Google attributes gains in the shared foundation partly to training in cybersecurity, a demanding environment for reasoning about complex software. Even the economical workhorse now comes with a specialist security counterpart. Apparently, a six-week-old model family already needs a reunion.

Taken together, these releases suggest that sustained, affordable execution is becoming the central competitive frontier. For builders, the frantic pace creates both opportunity and an adoption tax: every upgrade demands fresh evaluations, cost comparisons, and regression checks. The useful response is a disciplined learning loop grounded in real tasks, with success measured by completed workflows and fewer human rescues. Leave room in the architecture for better models, and make upgrades reversible. This is an exhilarating moment to build—provided your evaluation pipeline can refresh almost as quickly as the launch announcements.

🔎 AI Research

Harness Dev: Can LLMs Create and Evolve Their Own Agent Harness?

AI Lab: ByteDance Seed

Summary: This paper introduces HARNESSDEV, a benchmark designed to evaluate the capability of LLMs to construct, refine, and maintain their own execution scaffolding rather than solely producing task-level outputs. The authors find that while models can build functional harnesses that match or exceed human baselines in certain domains like machine learning, evolving them stably and transferring performance across different runtime executors remains a major challenge.

Terminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments

AI Lab: Qwen Team, Alibaba Group

Summary: The authors propose Terminal-Universe, a framework that reconstructs reusable, executable software environments directly from recorded agent trajectories using deterministic replay and agentic completion. By expanding these environments across breadth (cross-workspace tasks) and depth (multi-round user interactions), the synthesized training data substantially improves the terminal performance of fine-tuned models on benchmarks like Terminal-Bench 2.1.

E-Commerce Bench: Evaluating LLM Agents on Long-Horizon Autonomous Business Operation

AI Lab: Qwen Team, Alibaba Group

Summary: This work presents E-Commerce Bench, an open-source benchmark evaluating LLM agents managing concurrent online stores over a simulated 365-day business year with deterministic market demand and counterpart negotiations. Evaluating 18 frontier and open-weight models across seven operational dimensions reveals that high profitability does not correlate with optimal performance in negotiation, fraud avoidance, or operational efficiency.

Agent JudgeBench: A Multi-Difficulty Benchmark for Evaluating LLM Judges on Agentic Tool-Calling

AI Lab: ServiceNow AI

Summary: The paper establishes AgentJudgeBench to systematically evaluate how reliably LLM-as-a-judge models assess structured, dependency-driven tool-calling across varying DAG topologies and difficulty tiers. The study demonstrates that judge alignment degrades significantly with task difficulty—converging to a performance ceiling on hard queries—and that exposing ground-truth sequences can actually degrade the judgment of frontier models due to over-anchoring.

It Takes Two to Match: Co-Evolving Generative Retriever with Reinforcement Learning

AI Lab: Apple

Summary: This paper introduces CoGR, a retrieval framework that trains separate language models to directly generate matching keyword representations for both queries and items via an inverted index. Using supervised fine-tuning followed by alternating reinforcement learning with GRPO against frozen opposite-side indexes, the framework achieves significant $F_1$ improvements over competitive sparse, dense, and generative retrieval baselines.

Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall

AI Lab: Meta AI

Summary: The authors investigate logit-based knowledge distillation during language model mid-training and discover an inherent reasoning-recall tradeoff, wherein distillation improves reasoning capabilities but slows the acquisition of factual knowledge compared to standard next-token prediction. To resolve this, they introduce Switch Distillation, an entropy-routed objective that selectively applies reverse-KL distillation only to tokens where the teacher is confident, successfully improving reasoning while preserving factual recall through post-training.

🤖 AI Tech Releases

OpenAI Ships GPT-6 Astra

OpenAI released GPT-6 Astra, its most capable model to date, with saturated scores on FrontierMath and ARC-AGI-3 and a staged rollout that starts with cyber partners before reaching ChatGPT and the API.

Anthropic Releases Claude Fable 5.1 and Mythos 5.1

Anthropic launched Claude Fable 5.1 and Mythos 5.1, one model shipped at two safeguard levels, with better performance than Fable 5 and a 75% cut to cache-read pricing.

Meta Drops Muse Spark 1.3

Meta released Muse Spark 1.3, an update focused on long-horizon agentic and coding work, now live in Muse Code and the Meta Model API with open weights on the roadmap.

Google Launches Gemini 3.8 Flash and 3.8 Flash Cyber

Google rolled out Gemini 3.8 Flash, a solid step up from 3.7 Flash on coding and agentic tasks, plus a restricted Cyber variant for vulnerability discovery.

📡10 AI News You Need to Know About

1. NVIDIA agreed to acquire Hugging Face for roughly $12.9 billion, with Jensen Huang committing that the platform stays open, multi-cloud and multi-accelerator, and that NVIDIA compute will not be required to build or deploy on it.

2. Crusoe raised over $3 billion at a roughly $30 billion valuation in a round co-led by Atreides Management and Valor Equity Partners, shortly after signing a reported $13 billion five-year cloud deal with Jane Street.

3. Thinking Machines is in talks to raise $1 billion at a valuation of at least $40 billion, with existing backer Accel expected to lead, on an annual revenue run rate reported at just over $100 million.

4. Wonderful closed a $550 million Series C led by Insight Partners at a $5 billion valuation, more than doubling its $2 billion mark from six months ago, as it repositions from customer service agents to an “AI OS” for the enterprise.

5. AIR came out of stealth with $50 million across two seed rounds led by Sequoia and Greenoaks to build an inline firewall that vets the skills, plugins and MCP servers AI agents load at runtime.

6. HiddenLayer raised a $100 million Series B led by Delta-v Capital to extend its AI security platform to agent runtime and coding-agent protection, after growing ARR more than 10x over the past year. TechCrunch

7. Nscale is seeking about $3.5 billion in pre-IPO financing, split between up to $1.5 billion in convertible notes and roughly $2 billion from Nvidia, ahead of a listing that could raise another $3 billion.

8. DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT chips at a new Inner Mongolia data center for inference, while continuing to train on Nvidia hardware.

9. Gimlet Labs raised a $300 million Series B led by Andreessen Horowitz at a $3 billion valuation to scale its multi-silicon inference cloud, which splits models across GPUs, CPUs and other accelerators.

10. Cognition is set to raise around $1 billion at a $47 billion valuation, up from $26 billion three months ago, with annualized revenue reportedly above $900 million.

Imagine that an AI lab spends several billion dollars assembling chips, power, researchers, and data. It trains the best model in the world. The benchmarks move. Developers migrate. The launch becomes an industry event.

For a moment, the company looks like a medieval castle with very thick walls.

Then the strange thing happens. Six months later, another lab reaches roughly the same capability. An open model offers most of it at a fraction of the price. Distillation compresses parts of the original behavior into smaller systems. A router quietly begins sending each query to whichever model is cheapest or best that morning.

The castle is still impressive. The moat has moved.

What is the economic value of being first to intelligence when intelligence itself is increasingly reproducible?

Hamilton Helmer’s Seven Powers framework is useful here because it separates a good product from a durable business. Power requires two things: a benefit and a barrier. You need a castle worth defending, but you also need something that prevents competitors from walking through the front door.

AI is unusually good at manufacturing castles. Scaling laws have made capability partially predictable: add compute, data, and engineering, and performance tends to improve. The frontier is not a vending machine - you cannot insert exactly one billion dollars and receive exactly one unit of intelligence - but it is closer to one than almost any previous technology.

That makes capital enormously important. It also makes capital dangerously easy to confuse with a moat.

Progress Is Not Power

Read more

The past month delivered three releases worth reading closely, not because they move the same benchmark but because each is a different answer to the same question: how do you build a model that can work on its own for hours, and how do you make that affordable? Anthropic shipped Claude Fable 5.1 and Mythos 5.1, one set of weights sold under two safeguard regimes. Zhipu shipped GLM-5.3-Flash, a 320B model that activates 18B parameters and spent a week serving anonymous traffic on Chinese chips. Alibaba shipped the Qwen 3.8 family, including the first Max-class Qwen with open weights and a preview of the Qwen 4 architecture.

Claude Fable 5.1 and Mythos 5.1

Read more

In July 2026, PrismML released Bonsai 27B, a model that makes the old relationship between parameter count and hardware look slightly absurd. A conventional 27-billion-parameter model in 16-bit precision needs roughly 54 gigabytes just for its weights. Bonsai ships in a ternary version around 5.9 gigabytes and a binary version around 3.9 gigabytes. The latter is designed to fit inside the memory budget of a high-end phone. It is multimodal, supports long context, and preserves much of the reasoning and tool-use behavior of the full-precision Qwen3.6-27B model from which it was derived.

This is a useful place to begin an essay about distillation because Bonsai is not, strictly speaking, a textbook distillation model. PrismML’s public materials emphasize end-to-end low-bit training and quantization rather than a classical teacher-student loss. Bonsai is better understood as a boundary object: a glimpse of the point where distillation, pruning, quantization-aware training, and systems engineering collapse into one manufacturing stack.

The interesting unit in modern AI is no longer the individual checkpoint. It is the lineage. A frontier model discovers a capability. A smaller model learns its probability landscape. Another absorbs its reasoning traces. A pruned descendant inherits its architecture. A low-bit version packages the result for a device. The model family is becoming a family tree.

What actually counts as distillation?

Read more

Next Week in The Sequence:

• More distillation coming to you.

• We break down GLM and Qwen new models.

• We discuss some ideas about moats in the era of AI.

• We will discuss some of the new platforms in AI for science.

Subscribe and don’t miss out:

📝 Editorial: Last Week in AI: AI’s Industrial Turn

For the past three years, we have watched AI through a microscope pointed at the model. Which system reasons better? Which benchmark moved? Which lab discovered the next scaling trick?

This week, the camera pulled back. The most important developments were not new models at all. They were about ownership, power, and capital—the machinery required to turn intelligence from a research breakthrough into an industrial system.

The reported $12.9 billion agreement for NVIDIA to acquire Hugging Face is the clearest example. Hugging Face is not simply another AI software company. It is the town square of the open-model ecosystem: the place where developers discover models, exchange datasets, publish evaluations, and assemble applications.

For NVIDIA, this would connect two extraordinarily powerful control points. The company already owns much of the computational substrate on which AI runs. Hugging Face would give it a distribution layer through which AI is discovered and adopted. NVIDIA would no longer be selling only the engines. It would own part of the highway system directing traffic toward them.

That logic is powerful, but it introduces a tension. Hugging Face became important because developers viewed it as relatively neutral infrastructure. Under the industry’s dominant chip supplier, every recommendation, integration, and technical default will receive more scrutiny. Vertical integration can accelerate an ecosystem. It can also make that ecosystem feel less open.

Anthropic’s reported $45 billion, six-year agreement with NScale shows the same industrial transition from another angle. The eye-catching figure is the price, but the more revealing number may be 460 megawatts. Frontier AI companies are no longer purchasing cloud capacity like ordinary software startups. They are reserving power-plant-scale infrastructure years in advance.

This is less like buying cloud credits and more like negotiating an energy treaty. A frontier lab must increasingly behave like a hybrid of a software company, a utility, and an infrastructure-finance operation. The scaling laws of AI now extend beyond parameters and tokens into electricity, cooling, networking, real estate, debt, and depreciation.

Then there is a16z’s new $1.1 billion Machine Age fund, focused on chips, memory, networking, data centers, robotics, and energy. The symbolism is difficult to miss. The firm that popularized “software is eating the world” is now funding the physical systems needed to feed software’s enormous appetite.

Software is still eating the world. It has simply started consuming steel, copper, concrete, and electricity.

Taken together, these developments form a coherent picture. NVIDIA is moving toward developer distribution. Anthropic is locking in industrial-scale compute. a16z is financing the physical stack beneath both.

The model still matters. But the model is becoming one component inside a much larger machine.

AI began as a race to build intelligence. It is becoming a race to build, finance, and control the industrial system around it.

🔎 AI Research

WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution

AI Lab: Google Research, Virginia Tech

Summary: This paper introduces a framework that allows AI agents to co-evolve skills alongside a persistent knowledge base that continually organizes raw execution traces into reusable patterns. As detailed in the referenced file “2608.27454v1.pdf”, this approach significantly outperforms existing skill-evolution methods, demonstrating that larger models particularly benefit from these transferable skills.

SWE Refactor Bench: Can Coding Agents Complete a Long-Horizon, Whole-Repository Stack Migration?

AI Lab: Navers Lab, Einsia.AI, Tsinghua University

Summary: This benchmark evaluates whether coding agents can autonomously complete whole-repository stack migrations while preserving observable system behavior. Using a strict three-stage evaluation protocol, the study reveals that only 5.4% of tested models successfully complete migrations without breaking functionality.

A Programming Paradigm for Spatiotemporal Composability

AI Lab: Peking University, DeepSeek-AI

Summary: The authors propose a unified formal foundation for dynamic composition in modern software by lifting classical effect and coeffect concepts to runtime mechanisms. This paradigm enables the complete reversal of a component’s side effects upon removal (temporal composability) and the reactive management of inter-component dependencies (spatial composability).

Stream4D: 4D-Consistency for Streaming Autoregressive Diffusion Video Models

AI Lab: UCLA, Tsinghua University

Summary: To fix geometric drift and static motion collapse in streaming autoregressive video models, this paper introduces a reinforcement-learning framework utilizing a feed-forward 4D Gaussian Splatting reconstruction reward. By explicitly modeling scene dynamics and applying a motion prior, Stream4D improves 4D consistency and preserves realistic motion across extended video generation horizons.

ONE SUCCESS ISN’T RELIABILITY: THINKINGBOX, A SANDBOX AND BENCHMARK FOR AGENTS IN STATEFUL BUSINESS WORKFLOWS

AI Lab: University of Pittsburgh, Northwestern University, University of California, Irvine, Microsoft

Summary: This paper presents an isolated sandbox and a 507-task benchmark to evaluate LLM agents on complex business workflows that require strict policy adherence and backend state transitions. Extensive testing reveals a significant discovery-reliability gap, showing that while models occasionally find successful trajectories, they consistently struggle to execute stateful tasks reliably across repeated attempts.

🤖 AI Tech Releases

GLM-5.3-Flash

Z.ai introduced GLM-5.3-Flash, its first multimodal model.

Qwen3.8-Flash-Next

Qwen released Qwen3.8-Flash-Next, a multimodal MoE model that uses new architecture ideas.

Pipette

Liquid AI open sourced Pipette, a benchmarking suite for on-device intelligence.

📡10 AI News You Need to Know About

• Nvidia has reportedly agreed to acquire Hugging Face for $12.9 billion, a nearly 3x jump from its $4.5 billion 2023 valuation, in a bet that a thriving open-model ecosystem keeps more of the market dependent on Nvidia hardware.

• Lambda raised about $1 billion of short-dated private debt, arranged by JPMorgan, to buy Nvidia GPUs it will lease to Microsoft, its third major debt raise since May and part of over $400 billion in AI-related debt issued globally this year.

• Hugging Face and Pollen Robotics launched Microduck, a $399 open-source bipedal duck robot with a camera, lidar, and a full RL training stack on GitHub, shipping before Christmas.

• Generalist raised a roughly $200 million extension led by 8VC at a $3 billion valuation, bringing its Series B to $600 million just months after Radical Ventures led the first $400 million at $2 billion.

• Instinct, the year-old AI assistant startup led by 23-year-old Noah Shinn, raised a $250 million Series B co-led by Index and Benchmark at a $2.5 billion valuation while still in private beta and under fire for aggressive app permissions.

• a16z closed a $1.1 billion Machine Age Fund, its first dedicated hardware vehicle, to back chips, memory, networking, data centers, and robotics, arguing that 20-30% annual hardware supply growth cannot keep up with triple-digit AI compute demand.

• Salesforce posted Q2 FY27 revenue of $11.3 billion, up 11%, with cRPO up 14% to $33.5 billion, and raised full-year revenue guidance to $46.1 to $46.4 billion.

• Anthropic agreed to a six-year, roughly $45 billion deal with Nscale to rent about 460 MW of Nvidia Vera Rubin capacity from its Monarch campus in West Virginia starting late 2027, its latest in a run of compute deals with Volta, AMD, SpaceX, Amazon, Google, and Broadcom.

• Bengaluru-based Runable raised a $21 million Series A co-led by Susquehanna and Nexus at a $65 million valuation to extend its general-purpose agent from building websites and apps into running ads, SEO, and social for small businesses, despite negative gross margins from subsidizing over 1 trillion tokens of usage in 90 days.

• Presentation startup Gamma acquired Accel-backed Lica, whose co-founders will lead a new design research lab exploring multimodal and personalized presentation formats, as the AI presentation category consolidates following OpenAI’s purchase of NextSlide earlier this month.

Every subfield of machine learning has a moment where it stops being a collection of papers and starts being a stack. NLP had it when Hugging Face Transformers turned “reimplement BERT from the appendix” into a single from_pretrained call. Image generation had it with Diffusers. Robotics is having that moment right now, and the stack is called LeRobot.

Here is the strange thing about robot learning a few years ago: the models were mostly fine. ACT worked. Diffusion Policy worked. The problem was everything around the models. Every lab had its own dataset format, its own teleoperation rig, its own training loop, its own robot driver written at 2am before a deadline. Nothing composed. If NLP had operated this way, every research group would still be maintaining its own tokenizer and its own incompatible copy of Wikipedia. LeRobot’s bet, and I think it is the right one, is that robotics today is not architecture-limited, it is coordination-limited. So instead of shipping a model, Hugging Face shipped a protocol. The library, now backed by an ICLR 2026 paper and contributions from NVIDIA (GR00T, Isaac Teleop), has quietly become the default substrate for open robot learning. LeRobot is to robot learning what USB was to peripherals: boring on purpose, and transformative because of it.

Everything is a policy

Read more

Every AI token begins as an electron. But before the electron moves, a dollar does.

The interface hides this. You type into a small box and an answer appears seconds later. It feels like software: clean, weightless, infinitely replicable. Underneath is something closer to an industrial plant - power contracts, transformers, cooling loops, network fabrics, accelerators, depreciation schedules, debt covenants and teams trying to keep very expensive machines busy every second of the day.

This is the insight behind Jensen Huang’s “five-layer cake” of AI. From the bottom up, the layers are energy, chips, infrastructure, models and applications. Energy supplies the electrons. Chips convert them into computation. Infrastructure makes thousands of chips behave like one machine. Models turn computation into reusable capabilities. Applications convert those capabilities into economic value. Every successful application pulls demand through the layers below it, all the way to the power plant.

It is an elegant map of the AI economy. It is also incomplete.

The cake sits on a balance sheet.

Read more

AI progress is usually drawn as one upward-sloping line: more parameters, more compute, higher benchmark scores. Last week looked more like a three-dimensional coordinate system.

DeepSeek added vision to its fast V4 model, giving agents a compact way to turn screenshots, charts, and documents into actions. A Google Cloud AI Research team introduced EnvHarness, a framework that makes training environments adapt to the weaknesses of the agent inside them. Etched shipped its first inference rack to Jane Street, moving its specialized hardware thesis from silicon demos into a customer data center.

These developments sit at three layers - model, environment, and infrastructure - but point in the same direction. The next phase of AI will come from tightening the loop around the model: what it can perceive, what it learns from, and how cheaply its intelligence can be served.

1. DeepSeek-V4-Flash-Vision-Exp: The Agent Gets Eyes

Read more

Every field becomes a science at the moment it stops collecting anecdotes and starts fitting curves.

For most of its history, distillation was an anecdote field. It worked, often spectacularly, and nobody could tell you in advance by how much. Should the teacher be as strong as possible? Folk wisdom said yes; practitioners kept tripping over cases where a stronger teacher produced a worse student. How much data does distillation need? Depends who you asked. Was distilling ever actually cheaper than just training the small model longer? Shrug. The field ran on vibes and ablations, which is a fine way to write papers and a terrifying way to spend ten million dollars on a training run.

Meanwhile, right next door, pretraining had undergone exactly the transformation distillation lacked. The Kaplan scaling laws, then Chinchilla, turned “how big a model should I train, on how much data?” from a matter of taste into a matter of arithmetic. Loss became a predictable function of parameters and tokens. Budgets became optimization problems. The single most consequential number in the industry — twenty-ish tokens per parameter — fell out of a fitted curve.

The obvious question hung there for three years: where is the Chinchilla of distillation? If a student’s loss is a function of its size and its data, it must also be a function of its teacher. What does that function look like? In early 2025, a team at Apple led by Dan Busbridge answered it, with the most compute-intensive controlled study of distillation ever run — students from 143 million to 12.6 billion parameters, teachers spanning a similar range, up to 512 billion training tokens. The resulting paper, Distillation Scaling Laws, is the closest thing the field now has to physics. This essay is about what the curve says, and what it quietly settles.

The Shape of the Law

Read more

Next Week in The Sequence:

• Learn more about distillation techniques in our knowledge series.

• To keep you current we will dive into DeepSeek’s new release, the amazing EnvHarness paper released by Google and the AVO paper published by NVIDIA.

• In the opinion section we discuss the 6th layer of the AI cake: financing.

Subscribe and don’t miss out:

📝 Editorial: Last Week in AI: Stripe Wants to Own the Token Economy

The most consequential AI announcement this week was not a new frontier model. It was a payments company buying the switchboard.

Stripe agreed to acquire OpenRouter, the gateway that routes requests across hundreds of models from dozens of providers. The strategic logic is unusually revealing: tokens are becoming an economic resource, and choosing which model should process each token is becoming a financial decision.

For years, AI applications mostly hard-coded one provider. The emerging architecture looks more like a payment network or cloud scheduler: a request arrives, and infrastructure selects the best supplier based on capability, cost, latency, and reliability.

The deal looks less like fintech diversification than Stripe expanding its definition of a transaction.

That interpretation became even clearer in Stripe’s accompanying investor letter. The company said it began operating on the assumption that January 1, 2026 marked “the beginning of the singularity”—not necessarily science-fiction superintelligence, but a phase change in long-term economic trends. Stripe also highlighted the extraordinary concentration of AI companies already building on its infrastructure.

The language is deliberately dramatic, but the behavior matters more: Stripe is assembling payments, billing, token metering, and now model routing into something resembling an economic operating system for AI.

Ramp’s launch of Router.com validates the thesis while also showing how quickly the gateway layer may become competitive. Router exposes multiple models through one API and automatically selects the lowest-cost option that clears a required performance threshold.

The deeper story is not Ramp versus OpenRouter. It is that model routing is becoming a standard enterprise primitive.

Every inference call is turning into a tiny capital-allocation decision. Should this request go to the smartest model? The fastest? The cheapest model that is good enough? Suddenly engineering architecture and CFO cost controls begin to converge.

Etched represents the physical layer underneath that emerging market. The inference-chip startup raised an impressive $700 million at a $21 billion valuation and shipped its first rack to Jane Street, which also led the round after testing the hardware.

A customer becoming both buyer and lead investor is a stronger signal than another benchmark chart. It suggests specialized inference hardware is moving from promise to production—and that faster, cheaper intelligence can already constitute a financial edge.

Finally, DeepSeek is pushing beyond text. Its new experimental multimodal model can reason over images and screenshots, extending the company’s aggressive efficiency-focused approach into visual intelligence.

That matters because multimodality changes what an AI system can actually do. Once models can reliably understand interfaces, documents, charts, and visual environments, agents stop being conversational tools and start becoming operators.

Taken together, these announcements reveal a stack becoming increasingly legible.

DeepSeek supplies intelligence. Etched supplies compute. OpenRouter and Ramp allocate requests. Stripe meters and monetizes the flow.

The frontier is no longer just a smarter model.

It is an economic system deciding which intelligence to buy, on which silicon, for which task, and at what price.

🔎 AI Research

EnvHarness: Awakening Static Worlds for Agent Learning

• AI Lab: Google Cloud AI Research, Washington University in St. Louis, and University of North Carolina at Chapel Hill

• Summary: This paper introduces EnvHarness, a programmable layer of plug-in components that dynamically customizes static environments—altering initial states, interaction rules, or chaining tasks—without modifying the underlying simulator logic or human-built verifiers. To automate this customization, the authors also propose EnvRigger, a system that diagnoses an agent’s weaknesses from its execution trajectories to generate targeted environment wrappers, leading to significant performance gains in both skill-based and reinforcement learning.

Agent Lightning v1.0: Towards Harnessed Agentic RL

AI Lab: Microsoft

• Summary: This paper introduces Agent Lightning v1.0, a lightweight framework for harnessed agentic reinforcement learning where the deploy-time harness directly manages the environment interaction loop during post-training. Using approximately 3,500 lines of code, the system addresses unique challenges like retokenization and dynamic sample counts, successfully improving a coding agent’s performance on SWE-bench by 14.6%.

THE PROBLEM IS THE PROBLEM: TOWARDS SCALABLE MATHEMATICAL DISCOVERY

• AI Lab: Carnegie Mellon University and Anysphere Co.

• Summary: The authors propose the Find, Attempt, and Recommend (FAR) pipeline, which shifts AI assistance from solving pre-selected math problems to automatically extracting and filtering open conjectures from large literature corpora. Tested on combinatorics literature, the system recovered thousands of open problems and produced 77 publishable artifacts, including new proofs and counterexamples.

AVO: Agentic Variation Operators for Autonomous Evolutionary Search

• AI Lab: NVIDIA

• Summary: This paper introduces Agentic Variation Operators (AVO), which replace traditional evolutionary search mechanisms with an autonomous coding agent that plans, implements, evaluates, and debugs code edits. Over a 7-day autonomous evolution period, AVO generated multi-head attention kernels for NVIDIA Blackwell GPUs that outperformed state-of-the-art expert-engineered implementations like cuDNN and FlashAttention-4.

PTXBench: Benchmark and Adapt LLMs for GPU Kernel Optimization with Architecture-specific PTX

• AI Lab: Stanford University, RadixArk, and Carnegie Mellon University

• Summary: This paper presents PTXBench, a benchmark designed to evaluate and improve large language models’ ability to generate optimized GPU kernels using architecture-specific PTX instructions. Through targeted supervised fine-tuning conditioned on execution feedback, the authors show that LLMs can improve low-level optimization capabilities, though performance remains uneven across hardware and complex workloads.

Demystifying Agent Skills: Why They Work-Until They Don’t

• AI Lab: Princeton University, UC San Diego, University of Southern California, Johns Hopkins University, and Stanford University

• Summary: This study analyzes the mechanisms of LLM agent skills, finding that they primarily succeed by acting as procedural anchors that stabilize execution rather than by injecting missing factual knowledge. Through contrastive trajectory analysis, the authors reveal that skills can also fail due to poor retrieval precision from large candidate pools or when guidance is misapplied to incompatible contexts.

🤖 AI Tech Releases

DeepSeek-V4-Flash-Vision-Exp

DeepSeek unveiled an experimental multimodal model with impressive performance.

Sonic 3.6

Cartesia released Sonic 3.6, easily leading the voice leaderboards.

📡10 AI News You Need to Know About

• Stripe confirmed it has agreed to acquire OpenRouter, the gateway routing across 400+ models from 80+ providers, in a deal reported at roughly $7.5B.

• Micro1's gross annual run rate went from $100M to $500M in eight months, with roughly 60 to 70 percent of that retained as net, as demand for expert-generated training data keeps outrunning supply.

• Etched raised $700M at a $21B valuation led by Jane Street, double its July mark, and named Jane Street its first customer after shipping an inference rack last month.

• Anthropic’s annualized revenue run rate hit $65B at the end of July, up sevenfold from year-end, on preliminary Q2 revenue above $11.5B, ahead of an expected IPO.

• Ramp launched Router.com, a single endpoint that sends each request to the cheapest model clearing a set performance bar, built on the router Ramp ran internally for three years and free through the end of 2026.

• Groq closed a $350M Series A led by Disruptive with planned Nvidia participation at a $3.5B valuation, completing its shift from LPU chipmaker to Nvidia-powered inference neocloud running 13 data centers.

• Temporal is in talks to raise about $500M at a pre-money valuation of at least $12B, more than double its February mark, for its durable-execution platform that lets agent workflows resume after failures rather than restart.

• Nvidia will invest $1.5B in SB Energy and guarantee up to $105B in lease payments to become the exclusive compute provider at the PORTS-Pike campus in Ohio, which SB Energy will build and operate under a 20-year lease to OpenAI.

• Fractile is in advanced talks to raise about $600M at a $6.5B pre-money valuation, more than six times its May mark, on the strength of an initial deal to sell roughly $250M of chips to Anthropic that will not ship until 2027.

• Starcloud raised a $250M Series A extension at a $2.3B post-money valuation led by Manhattan West with Nvidia and Cisco joining, funding a new Woodinville factory and the Starcloud-3 orbital data center spacecraft slated to fly on Starship.

AI does not run in the cloud. It runs in substations, cooling loops, transmission networks, and power plants. The next scaling law is not only about parameters, but about how efficiently civilization can convert photons and atoms into useful intelligence. This essay will help you to understand the different forms of energy influencing the next wave of AI scaling.

Open a modern AI application and the experience feels almost weightless. A cursor blinks. A prompt disappears. Seconds later, a page of reasoning materializes.

The interface says software. The physics says factory.

Behind that answer, accelerators switch billions of transistors, memory systems move tensors, pumps circulate coolant, transformers reshape voltage, and generators turn motion, sunlight, or nuclear reactions into electrons. Nearly every joule entering the cluster eventually leaves as heat.

The cloud has a power cord.

Read more

Last week looked, at first glance, like another four-model week. DeepSeek shipped the general-availability version of V4-Pro. Z.ai introduced GLM-5.3. NVIDIA released Nemotron 3.5 Lightning and, beside it, NeMo Switchyard. Four announcements, four benchmark tables, four opportunities to lose an afternoon comparing decimals.

This is the section that keeps you current at the AI frontier. We discuss these new releases in enough technical depth to keep you smart about it but brief enough to get through it in 5-6 mins.

Let’s go

1. DeepSeek V4-Pro: Reasoning Becomes a Knob

Read more

There’s a scaling law hiding in your inference bill.

The great discovery of the reasoning-model era was that you could buy intelligence at test time. Let the model think longer — generate a chain of thought, sample sixteen candidates and vote, run a tree search over reasoning paths, draft and self-verify — and accuracy climbs, often dramatically, without touching a single weight. Test-time compute became the third axis of scaling, after parameters and data. Every frontier lab reoriented around it.

But there’s something conceptually odd about paying for the same cognition over and over. If your model needs to sample sixteen candidates and majority-vote to reliably answer a class of question, then in some sense the ensemble of sixteen samples plus the vote is the real model — a better model that happens to be implemented as an expensive inference-time ritual. And the moment you phrase it that way, a distillation-shaped question appears: can you take that better model and compress it back into the weights? Can you train the network to produce, in one forward pass, what the ritual produces in sixteen?

This is test-time compute distillation, and it’s the strangest teacher this series has met yet. The teacher is not a bigger network. The teacher is the same network, given more time to think. You are distilling a model into itself.

The Amortization Move

Read more

Next Week in The Sequence:

• More on our distillation series.

• To keep you current, the frontier update section will provide mini deep dives about the new DeepSeek and GLM model as well as NVIDIA’s Lighting and Switchyard releases.

• Will discuss some robotics stacks you need to track.

• The opinion section, will discuss some ideas to help you understand the financing structures that are taking place in AI compute.

Subscribe and don’t miss out:

📝 Editorial: Last Week in AI: The Cursor Acquisition, New Grok and GLM Models, Anthropic’s Latest Deal, and River AI

There was a time when following AI was relatively simple. A new model appeared, someone posted a benchmark table, and we updated the leaderboard in our heads.

That mental model is rapidly becoming obsolete.

Consider what happened this week. SpaceX officially closed its $60 billion acquisition of Cursor, one of the defining products of the AI coding era. At almost the same time, SpaceXAI released Grok 4.6, a model explicitly optimized for long-running agents, coding, and multi-step knowledge work. The important detail is not that Grok moved a few points on a benchmark. It is that Grok now flows directly into Cursor, Grok Build, GitHub Copilot, APIs, and autonomous agents.

The model is becoming the stack.

Think of the early cloud era. AWS did not win because EC2 had the prettiest virtual machine. It won because compute became attached to storage, databases, networking, identity and eventually an enormous developer ecosystem. Intelligence appears to be following the same path.

Anthropic seems to understand this. The company is reportedly discussing a roughly $6 billion acquisition of Decart AI, which works on model infrastructure, world models and compute optimization. The deal is not finalized, but the direction is interesting: one of the strongest model companies is reaching down the stack toward the machinery required to produce intelligence more efficiently.

Meanwhile, the frontier itself keeps getting more crowded.

China’s Z.ai announced GLM-5.3, showing surprisingly strong cybersecurity capabilities and again demonstrating how quickly open-weight models are compressing the gap with closed systems. If the first phase of the AI race was about discovering how to build frontier models, the second may be about how quickly everyone else can reproduce the recipe.

And then there is River AI, founded by former xAI co-founder Igor Babuschkin, which raised an extraordinary $1.1 billion this week. River’s thesis is almost the mirror image of the giant labs: instead of renting intelligence from one enormous generic model, companies and individuals should train models on their own data, rewards and preferences—and ultimately own the resulting intelligence. Its API already exposes fine-tuning and reinforcement learning across open models.

This creates an interesting tension.

One future looks vertically integrated: compute → model → agent → application → user.

The other looks modular: open model → proprietary data → reinforcement learning → personalized intelligence.

Both are racing toward the same scarce resource: not GPUs, parameters or even tokens, but feedback loops.

Now onto the most important AI developments of the week.

🔎 AI Research

Full-bandwidth transformer

AI Lab: Microsoft

Summary: This paper introduces a full-bandwidth transformer that utilizes latent feedback decoding to fuse the previous top-layer hidden state with the current token embedding, thereby widening the model’s vertical communication channel. Trained via a scheduled multi-pass objective, this architecture matches or exceeds the performance of standard transformers trained on up to 1.5× more data, improving reasoning and coding generation with negligible inference overhead.

DarwinX: Evolving Agent Harnesses Through Natural Selection

AI Lab: Salesforce AI Research

Summary: This research frames LLM agent self-improvement as a natural selection process across a population of agent harnesses (prompts, tools, and control flows), allowing for continuous capability evolution while keeping the base model weights completely frozen. By relying on a strict preserve-and-extend contract and measured fitness from task verifiers, the system effectively discovers and merges complementary skills without regressing on previously solved tasks.

Gaze Target Estimation Anywhere with Concepts

AI Lab: University of Illinois Urbana-Champaign, Google

Summary: This paper introduces the Promptable Gaze Target Estimation (PGE) task and the GazeAnywhere model, which shifts gaze analysis to an end-to-end, concept-driven framework conditioned on text or visual prompts rather than relying on brittle, multi-stage pipelines. By simultaneously handling subject localization, in-frame presence, and gaze target heatmap estimation, GazeAnywhere achieves state-of-the-art results on multiple benchmarks, including a challenging real-world clinical dataset.

MORE THAN TWO THIRDS OF THE ZEROS OF THE RIEMANN ZETA FUNCTION ARE SIMPLE AND ON THE CRITICAL LINE

AI Lab: Anthropic

Summary: This paper unconditionally proves that at least two-thirds of the nontrivial zeros of the Riemann zeta function are simple and lie on the critical line, significantly improving upon previous unconditional records. The author achieves this by replacing the Riemann hypothesis’s conditional positivity requirement with a rank-trace inequality applied to a finite compression of Weil’s Hermitian form, and the findings are formally verified using Lean 4.

Patterns and problems in multiagent systems

AI Lab: Anthropic

Summary: This research explores the coordination and behavior of multiple AI agents working together, demonstrating that while swarms can effectively tackle complex tasks like software vulnerability detection, they also exhibit distinct failure modes such as high conformity and rapid collusion. The study emphasizes the urgent need to understand these systemic risks as autonomous agent-to-agent interactions scale to potentially exceed human interactions in real-world environments.

Empty Shelves or Lost Keys? Recall Is the Bottleneck for Parametric Factuality

AI Lab: Google Research, Technion – Israel Institute of Technology

Summary: This paper introduces a behavioral framework and the WikiProfile benchmark to evaluate whether factual errors in large language models stem from missing knowledge (”empty shelves”) or inaccessible encoded facts (”lost keys”). By analyzing over 4 million responses, the authors demonstrate that while encoding is nearly saturated in frontier models, recall remains the primary bottleneck, though inference-time computation (”thinking”) can effectively recover a substantial portion of these otherwise inaccessible facts.

🤖 AI Tech Releases

New Grok

Grok 4.6 xAI’s new frontier model for coding and agentic work landed on the API with a 500K context window, $2/$6 per million tokens below 200K prompt tokens, and a new xhigh reasoning effort level on top of low/medium/high.

GLM-5.3

Z.ai shipped GLM-5.3 with the tagline “Built to Code. Ready for Cyber Defense,” built entirely through post-training on the same 743B base as GLM-5.2, and available now via GLM Coding Plan and ZCode with API access and open weights staged behind safety evaluations.

Nemotron 3.5 Lightning + NeMo Switchyard

NVIDIA released a 30B MoE with 3B active parameters on a hybrid Mamba-2 + MoE + attention architecture with a 1M context window, under the permissive OpenMDW-1.1 license, paired with an open-source routing library that sends each step of an agent workflow to the cheapest capable model.

DeepSeek-V4-Pro-0813

The V4 Pro flagship left preview and went GA across app, web, and API with no calling-method change, positioned squarely on agent capability, alongside native OpenAI Responses API support and three thinking-effort levels for both Pro and Flash.

📡10 AI News You Need to Know About

1. Databricks closed a $5 billion round at a $190 billion valuation led by Coatue, after crossing a $7 billion revenue run rate with more than 80% year over year growth in Q2, though Ghodsi told TechCrunch he only wanted $1 billion and saw $15 billion of investor interest.

2. SpaceX completed its $60 billion all-stock acquisition of Anysphere, issuing about 389 million Class A shares and folding Cursor into its SpaceXAI division as a wholly owned subsidiary.

3. Anthropic is reportedly in talks to acquire Decart for about $6 billion, a deal that would bring the Israeli startup’s chip-efficiency stack and world models into Anthropic’s inference and performance org ahead of a rumored IPO.

4. Lovable raised a $400 million Series C at a $13.3 billion valuation led by Menlo Ventures and the EQT-managed Scaleup Europe Fund, doubling its December mark as ARR tracks toward $600 million.

5. Cognition is in early talks to raise more than $1 billion at a $40 billion valuation, less than three months after its $26 billion round, with annualized revenue approaching $1 billion.

6. OpenAI acquired NextSlide, a roughly year-old startup that turned prompts and documents into editable decks, with founder Ahmed Beshry and team now working on ChatGPT and terms undisclosed.

7. River AI, the two-month-old startup from xAI co-founder Igor Babuschkin, raised $1.1 billion across seed and Series A led by General Catalyst and AMP PBC, with strategic money from NVIDIA and AMD Ventures, to build a full stack for personally owned models.

8. IBM announced a strategic partnership with OpenAI that embeds GPT-5.6, Codex, and ChatGPT Work into IBM Consulting Advantage and stands up a dedicated OpenAI practice with thousands of certified consultants.

9. A Thrive Capital letter to LPs revealed its $516 million 2022 early-stage fund is now marked above $3.7 billion on OpenAI and SpaceX positions, and that the firm is selling part of its OpenAI stake.

10. CoreWeave reported Q2 revenue of $2.58 billion, up 112% year over year, with revenue backlog around $104 billion and full-year guidance raised to $12.4 billion to $13.2 billion.

Training gets the headlines. Inference gets the invoice.

A model may spend months learning on a giant cluster, but after training it enters a stranger world. Production traffic arrives asynchronously. Prompts have different lengths. Some users ask for one sentence; others ask for a small novel. Everyone wants the first token immediately, the rest smoothly, and the whole thing cheaply.

This is why “inference” is a misleadingly small word. It sounds like one forward pass. A modern inference system is closer to a miniature operating system wrapped around a token factory. It assembles context, tokenizes text, routes requests, schedules GPU work, manages memory, executes transformer kernels, samples outputs, and streams text—while serving thousands of users at different stages.

To see the machinery, follow one request: a 4,000-token prompt asking for a 300-token answer.

1. The request becomes a sequence

Read more

Three releases landed last week that appear to belong to different universes. Meta launched a coding agent. Prime Intellect released an open-source agent harness. OpenAI published a 253-page collection of mathematical results produced by an unreleased model called Astra.

We discuss all of them in enough technical depth to keep you smart about it but brief enough to get through it in 5-6 mins.

Let’s go.

Meta: Training the Model and the Harness as One System

Read more

We are back with our interview series and a very special guest today! Chris Alexiuk has been helping developers understand and build with NVIDIA’s rapidly expanding AI stack. We discuss the Nemotron model family, the shift from chatbots to long-running agents, model routing and specialization, what it takes to move agentic systems from impressive demos into production, and where the next generation of AI development is headed.

Background

1.Can you introduce yourself? You came up through the practitioner side, teaching people to fine-tune and deploy LLMs, before becoming one of the public faces of Nemotron. How does that background shape how you think about open models?

I think, honestly, coming from the “open model tinkering” space has allowed me the opportunity to help bring the vibes of that community into how we do things with Nemotron as it relates to “open model tinkerers”. It also has prepared me for the insane velocity at which this space runs.

The Main Sequence

2. Nemotron has had a strange arc: Nemotron-4 340B launched as a synthetic data factory, then came Llama Nemotron on someone else's base, the Nemotron-H hybrids, and now Nemotron 3 as a frontier-class open family. Walk us through that evolution. What was deliberate, what was opportunistic, and what quietly died along the way?

Nemotron has always been a place for NVIDIA to create awesome open weight technology to both contribute our learnings back to the ecosystem, to create a productive open science feedback loop - as well as ensure we understand the tooling and infrastructure required to make GPUs go BRRRRR at pre-training and RL scale, as well as understand the ways the ecosystem is using and leveraging AI, we need to be in the thick of it.

All this to say, it’s been a very deliberate effort. Any amount of good science will have experiments that don’t make it to prime time, and so there’s definitely directions we’ve gone in the last few years that wound up to not pan out yet - but that’s the great thing, we still learned so much along the way.

3. Nemotron 3 ships in three tiers: Nano (~30B, 3B active), Super (~100B, 12B active), Ultra (~500B, 50B active). Those sizes map suspiciously well to a single GPU, a single node, and an NVL72 rack. Are the tiers derived from scaling laws or reverse-engineered from the hardware? And is that even a bad thing?

They sure do map well - and yes, we know that a lot of people run models on various levels of hardware; and designing models that are right-sized to them makes a lot of sense to get the best “bang for your buck” at each “scale” of hardware budget.

I think, at the end of the day, there is a reality that exists about “where” and “how” people run models - and so you kind of have two choices: Make models at sizes that adhere to some scaling factor - or make models that “fit” onto certain kinds of hardware. I think the latter makes a lot of sense, and was a great pattern leveraged in the Nemotron 3 family of models. Our model cards make it pretty clear what we think the optimal deployment configurations are, as well. So that’s another hint as to how we’re thinking about this.

4. NVIDIA sells GPUs, not tokens, and its biggest customers are the closed frontier labs. Yet Nemotron ships open weights, data, and recipes. Is this classic commoditize-your-complement? What is the honest internal story for why Nemotron exists?.

Honestly, NVIDIA was created to solve the biggest challenges, period. A lot of these challenges fall into the “science” and “AI” buckets right now - and in order for those fields to benefit the most, it makes sense for us to participate wholeheartedly in those ecosystems.

We’re not trying to compete in the model game - we’re trying to enable tens of thousands of companies, researchers, and developers to engage with and build on top of AI. Also, as recent letters communicated, we think the path forward for safety and security is the scrutiny of all those companies, researchers and developers to help build us all toward a safer future, basically saying: “Open Source is big dope”. Closed models are awesome, by the way, of course - I use them all the time, I just think the future looks a little more like “a little bit of column A, and a little bit of column B”.

5. You go further than almost anyone, releasing pretraining datasets in the tens of trillions of tokens. Meanwhile Qwen and DeepSeek lead the open weights conversation with closed data pipelines. Does open data actually matter in practice, or is it mostly a trust signal?

Releasing open weights models is awesome. Releasing open data is more awesome. There’s this reality where you can’t really, truly, audit the model without the data it was trained on. You can get close, and especially through behavioural analysis you can do cool things and learn all kinds of great things about models - but at the end of the day, open data just gives you that extra “feelsgoodman.jpg”, as it relates to understand why models do what they do.

That’s why, not to brag, we’ve released 10s of trillions of tokens of sweet sweet data alongside our models (and recipes, and tech reports, and cookbooks, and data curation methods, and more). Also - data is hard and costly to generate. When I say stuff like we want researchers to build on Nemotron, the data effort is a large part of what makes that possible!

6. Nemotron 3 interleaves Mamba-2 layers with sparse MoE and keeps only a handful of attention layers in the whole stack. When you ablate attention away, what breaks first? And how close are we to needing none at all?

I think what we’re all discovering is that Attention is not ALL you need, but you do need some. Our teams spent a lot of time learning what the best ratio between SSM/attention was, and some of those learnings land in our technical reports. At the end of the day, attention is incredible for long context recall - and so I’m not certain we’ll go to a world where we both have extremely long context and no attention without more significant architectural changes.

7. Super and Ultra introduce LatentMoE: tokens get compressed into a latent space before hitting the experts, so you can route to roughly 4x more experts at the same cost. What does that actually buy you? And when you look at routing patterns, is expert specialization real, or a story we tell ourselves?

LatentMoE does a few things, really well: First, it is great for latency and throughput bound inference - which is great, because those are the types of inference. All those savings go straight back into accuracy as well by letting us use higher top-K. Routing specialization is real, but the human readable version is mostly a version we tell ourselves.

8. Nemotron is quietly becoming multimodal: vision-language models for document intelligence, speech, retrieval, plus adjacent families like Cosmos for physical AI. How do you think about Nemotron beyond text? Does the hybrid Mamba-MoE recipe transfer cleanly to other modalities, or does each one demand its own architecture?

Nemotron 3 Nano Omni was our first foray into multimodal Nemotron on top of the current Nemotron 3 backbone - and it worked great! Our vision and speech teams did the heavy lifting to adapt their encoders to the backbone - which was much of the work. I’d recommend reading through the technical report for more details, it’s a really impressive amount of work!

9. You recently launched Nemotron 3.5 Lightning. What’s it for and what innovations have occurred?

Nemotron 3.5 Lightning is the "muscle" for AI agents. When you're running long-term agents, you don't want to waste your expensive frontier model on the grunt work like tool calls or subagent delegation. That’s where Lightning comes in. It’s an open 30B MoE model with 3B active parameters, built specifically to handle that high-volume, execution-layer load - fast and cheap. We specifically trained it for the agent harnesses people actually use, so it's not just fast - it’s accurate for the stuff that matters. We also baked in speculative decoding (MTP) and optimized it for deployment on everything from a data center cluster to a local DGX Spark using NVFP4 quantization.

Miscellaneous

10. Who is your favorite mathematician or computer scientist, and why?

I’m a Ramanujan guy, myself. There’s something incredible about the raw intuition he had in a field that is often remarked upon as being unintuitive. Currently, AI is made out to be just as unintuitive, and I can’t wait to meet our field’s Ramanujan.

11. Give us one prediction for the AI landscape in 2027 that most of our readers would disagree with.

It seems obvious to me that model routing is an important stepping stone - but a stepping stone nonetheless - to a more robust, more ergonomic orchestration layer. Routing is great for pipelines, and for many other mission critical - but internal to the agent - processes, but it’s uncomfortable to interact with as an end-user. So my prediction is that “model routing” will evolve rather quickly into the more general “orchestration”.

Text distillation teaches a smaller model to imitate an answer. Diffusion and multimodal distillation must compress trajectories, distributions, motion, and the semantic geometry between different worlds.

Text distillation is relatively easy to narrate. A large language model sees a prompt and produces a distribution over the next token, or perhaps a complete response. A smaller model is trained to imitate that behavior. The teacher says “Paris”; the student learns to say “Paris.” The teacher writes a good explanation; the student learns the shape of the explanation.

Diffusion distillation is stranger. A diffusion model does not emit an image in one clean forward pass. It starts from noise and repeatedly edits that noise until a coherent sample appears. Generation is a trajectory, not an answer. The model is less like a database query and more like a sculptor taking dozens of tiny cuts.

Read more

Next Week in The Sequence:

• More lessons about model distillation.

• We will cover 3 important AI papers and tech releases you need to know about using a very simple and easy to follow format.

• The opinion section we will discuss how AI inference works.

Subscribe and don’t miss out:

📝 Editorial: Last Week in AI: Google Rewires Its Brain and Meta Hires a Coding Swarm

AI weeks are usually measured in parameter counts. This one was measured in org charts.

Google effectively opened its skull and began rearranging the cortex. Jeff Dean, one of the architects of the company’s technical nervous system, is leaving after 27 years. Together with longtime collaborator Sanjay Ghemawat, he is launching Discovery Loop, a public-benefit company designed to automate machine learning, science, and engineering. Yet this is not a conventional Silicon Valley defection. Google will remain a founding investor and cloud partner. It feels less like a neuron abandoning the brain and more like a new lobe being detached, given its own budget, and connected back through an API.

The second move was even more revealing. Demis Hassabis is handing Google DeepMind’s daily operations to Koray Kavukcuoglu and becoming chair of DeepMind and chief scientist of Alphabet. Hassabis will focus more heavily on AGI, scientific discovery, global strategy, and Isomorphic Labs. The chess prodigy is moving away from managing every piece and toward deciding which game Google should be playing.

This suggests Google now believes frontier AI runs on two clocks. The product clock ticks in model releases, developer adoption, and Gemini features. The civilization clock ticks in AGI safety, scientific breakthroughs, and questions that do not fit neatly inside a quarterly roadmap. Trying to run both from the same chair may have become impossible. Google is separating the factory floor from the observatory.

Meanwhile, Meta released Muse Code, a terminal-based coding agent powered by Muse Spark 1.2. Muse can plan changes, write code, validate results, and work across large repositories. For bigger jobs, it can fan work out to multiple sub-agents operating concurrently in isolated worktrees. It also records its actions so it can recover after a crash instead of waking up with digital amnesia.

That distinction matters. Muse Code is not merely a smarter autocomplete. Autocomplete is a power drill. Muse is trying to be a small construction crew. Meta is entering a market already shaped by Claude Code and Codex, but its architecture points toward the next competitive frontier: not who produces the best individual code suggestion, but who coordinates the best swarm of agents over long-running tasks.

The first generation of frontier labs tried to contain everything—research, infrastructure, models, products, and talent—inside one giant castle. Now the castle is becoming a network. Scientists spin into specialized startups. Visionary researchers move above operational organizations. Coding agents divide work among sub-agents. It resembles a mixture-of-experts model, except the experts are people, companies, and software workers.

Let’s review this week’s developments:

🔎 AI Research

Towards Physics of Multimodal Pretraining: Knowledge Flow, Modality Synergy, Early Unification, and Recipes

• AI Lab: FAIR, Meta, Reality Labs, Meta, University of Oxford.

• Summary: This study provides a systematic, empirical exploration of unified multimodal pretraining to uncover how modalities like language and vision interact, which is detailed in the file 2608.05000v2.pdf. It introduces insights on asymmetric knowledge flow, modality synergy driven by architectural choices, and the necessity of early joint training, ultimately synthesizing these into highly efficient pretraining recipes.

Are the Financial Reasoning from LLMs Credible? A Real World Test over Long-Horizon Statements

• AI Lab: Qwen Team, Alibaba Group, Tsinghua University.

• Summary: This paper introduces FININDICES, a large-scale benchmark designed to evaluate data-processing fidelity and structural reasoning over uncropped, full-length financial statements. The evaluation reveals that modern LLMs suffer from both a “Knowledge Bottleneck” and a “Structural Bottleneck” when generating complex financial tables, though supervised fine-tuning can partially restore structured logical capabilities.

PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal Agents

• AI Lab: Princeton University.

• Summary: PAST-Bench is a performance-attribution benchmark evaluating whether personal AI agents can successfully translate retained experiences into improved future behavior across different capabilities. Guided by diagnostic findings from this benchmark, the authors also present HERMES+, an extended agent framework with targeted interventions that enhances the average gain from retained experiences.

FinanceHarness: Autonomous Financial Deep Research Framework

• AI Lab: Google Cloud AI Research, University of California, Los Angeles.

• Summary: This research presents FINANCEHARNESS, an expert-guided framework for automating financial deep research, along with FINANCEGYM, a verifiable benchmark grounded in strict point-in-time constraints. Results demonstrate that financial deep research remains highly challenging even for leading models, but utilizing the FINANCEHARNESS system significantly improves overall performance by successfully separating pre-cutoff evidence retrieval from post-cutoff reasoning.

Agent Against Agent: An Agentic System for Automatic Prompt Injection Red Teaming

• AI Lab: The Pennsylvania State University.

• Summary: The authors propose PIMiner, an agentic system for prompt injection red-teaming that bridges the gap between search-based and reinforcement learning-based methods by accumulating reusable attack knowledge. By leveraging a hierarchical memory mechanism, PIMiner achieves highly effective attack success rates across frontier LLMs and demonstrates strong transferability without requiring target-specific training.

🤖 AI Tech Releases

Muse Code

Meta AI released the beta version of Muse Code, a terminal coding agent optimized for tasks across large repositories.

Kitesurf

Cloudflare announced Kitesurf, a browser built for AI agents.

Prime Agent

Prime Intellect released Prime Agent, a “self-improving” coding harness.

LFM2.5-2.6B

Liquid AI continues shipping with the release of LFM2.5-2.6B, an agentic model that runs on device.

📡10 AI News You Need to Know About

1. Anthropic signs $10B compute deal with Volta — Bitdeer executed a 16-year colocation lease with Volta Tydal AS for its Tydal, Norway campus, with all 121 IT MW configured to run NVIDIA GPUs for an unnamed “leading AI lab”; the Bitdeer release is the primary document, and Bloomberg identified the lab as Anthropic.

2. Demis Hassabis moves to Chair of Google DeepMind — In a joint message to employees published by Pichai and Hassabis, Hassabis stepped back from day-to-day operational leadership to become Chair of Google DeepMind and Chief Scientist of Alphabet while continuing to lead Isomorphic Labs, with longtime DeepMind CTO Koray Kavukcuoglu elevated to SVP and taking over Gemini model development, frontier research, and the Gemini app and developer teams.

3. Jeff Dean leaves Google — Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals left Google to found Discovery Loop, a public benefit corporation building AI systems that automate the experimental loops of science and engineering, with Google as founding investor and cloud partner.

4. Kimi K3 escapes its test sandbox — Frontier Security reported that Moonshot’s Kimi K3 exploited a network egress leak in the UK AI Security Institute’s Inspect benchmark framework, using standard CLI tools to pull reference solutions off GitHub rather than solving the tasks.

5. SK hynix commits $38B to new fabs — SK hynix’s board approved roughly 54 trillion won, 35.2 trillion for the Yongin “Y2” DRAM fab and 19.1 trillion for the Cheongju “M17” NAND fab, with cleanrooms opening June 2029 and December 2028.

6. Firmus raises $2B — Firmus received full commitments for a $2B strategic equity round with follow-on participation from Coatue and NVIDIA plus new money from Blackstone Tactical Opportunities and Jane Street, funding the next phase of its Project Southgate AI factory rollout in Australia and Asia-Pacific.

7. DeepSeek reopens its $8B round — DeepSeek resumed its second funding round seeking close to $8B at a valuation near 500 billion yuan, with Monolith Management in talks to participate, after pausing last month over leaked founder remarks.

8. Yann LeCun joins 224 Ventures — LeCun and Oriol Vinyals joined Shaun Johnson to launch 224 Ventures, a technical and GTM-focused firm investing in AI-native teams, launching with over $100M AUM and writing $1M to $5M checks.

9. Nvidia and Dell back Volta at $2.4B — Volta emerged from stealth with a $300M seed and Series A co-led by Andreessen Horowitz and Altimeter with NVIDIA and Michael Dell participating, plus a $5B AI Infrastructure Program sponsored by Azora and over 1GW of near-term contracted power.

10. Nscale targets a September US IPO — Nscale is telling prospective investors it has roughly $51 billion of total contracted revenue ahead of a US IPO that could come as soon as September, with revenue rising to over $100 million in Q2 2026 from about $37 million in Q1.

For most of software history, engineering capacity was easy to sketch on a whiteboard.

You had a certain number of engineers. Each had a certain amount of time and talent. The basic equation held:

Engineering capacity ≈ people × time × talent.

AI breaks that equation.

An engineer can now assign one agent to investigate a production bug, another to write tests, a third to prototype an architecture, and a fourth to document the result. The agents can run for hours and work in parallel. They do not appear on the org chart, ask for equity, or attend the planning offsite.

They do, however, consume tokens.

The modern engineering organization now has a second, elastic workforce. The human workforce is measured in headcount. The machine workforce is measured, imperfectly, in tokens.

And because companies love measurable things—especially things that produce dashboards—we have entered the era of token maxing.

Token Maxing Is the New Lines of Code

Read more