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

The demo carrying this entire release is boring on purpose. You ask Apptronik’s Apollo 2 to put the watering can into the green bin on the bottom shelf. It walks to the table, picks up the can, takes a few steps to the shelves, and places it where you asked.

Nothing in that sentence sounds hard until you remember what the previous Gemini Robotics models actually were. They were a torso bolted to a fixed base doing tabletop work. If there were legs, they belonged to somebody else’s controller.

Walking is not the interesting part. Boston Dynamics solved walking years ago with hand-tuned controllers and a lot of hydraulics. The interesting part is that the walking and the grasping came out of the same policy, conditioned on the same language instruction. Locomotion stopped being a subsystem you call and became part of the action space you predict. That is the real news in Gemini Robotics 2, and it is a bigger deal than the b-roll makes it look.

Three models, two access tiers

Read more

Every form of distillation in this series so far has quietly preserved one thing: teacher and student spoke the same dialect. A small transformer learned from a big transformer. The student was a compressed copy, then a more capable apprentice, then a reasoner trained on traces — but underneath, it was always the same kind of machine, attention layers stacked on attention layers, differing only in size.

Cross-architecture distillation breaks the last shared assumption. Here the teacher is a transformer and the student is not — it’s a state-space model, or a linear RNN, or some gated recurrent thing that has never computed an attention matrix in its life. You take a fully trained transformer and pour its capability into a fundamentally different computational substrate, and somehow the capability survives the transplant. The first time you see it work, it feels a little illicit, like recovering a person’s memories after swapping out their brain for different hardware.

This is the strangest corner of the distillation world, and also one of the most economically loaded. So it’s worth understanding why anyone would attempt something this perverse — and why, against reasonable expectations, it works.

The Arbitrage

Read more

Next Week in The Sequence:

• We continue our series about distillation with an awesome new technique.

• The AI of the week dives into Gemini Robotics 2.

• We will have a new section about AI in space.

• The opinion section get into a crazy idea for engineering teams in the era of tokens

Subscribe and don’t miss out:

📝 Editorial: Last Week in AI: Open Models, Intelligent Robots, and the Price of Conviction

AI spent this week speaking four languages: policy, models, robots, and markets. Strangely, all four delivered the same message. The AI race is moving beyond spectacular demonstrations and toward harder questions about distribution, embodiment, ownership, and economic returns.

Jensen Huang helped frame the policy debate by backing an industry letter defending open-weight models. This was not simply an argument about research culture. It was industrial strategy. The letter’s central idea is that American leadership cannot depend only on a handful of closed systems. It also requires an ecosystem in which startups, universities, enterprises, and public institutions can inspect, adapt, and operate advanced models themselves.

The timing was almost too perfect. Moonshot then released the weights for Kimi K3, a massive mixture-of-experts model with native multimodality and a one-million-token context window. K3 may not decisively surpass the strongest proprietary systems, but that is almost beside the point. Open models are no longer the minor leagues. They are becoming a parallel frontier—and Chinese laboratories are increasingly setting its pace.

Google DeepMind pushed the frontier in a different direction with Gemini Robotics 2. The release extends Gemini from understanding the digital world to controlling the physical one: planning complex tasks, coordinating full-body humanoid movement, and manipulating objects with greater dexterity. The deeper significance is architectural. The next model race may not be won by the system that writes the best answer, but by the one that can turn reasoning into reliable action. Robotics is where tokens acquire consequences.

Then markets supplied the warning label. Leopold Aschenbrenner’s Situational Awareness fund suffered a dramatic collapse and forced unwind after highly concentrated AI positions moved against it. The episode does not invalidate the long-term AI thesis. It illustrates something more uncomfortable: a secular prediction can be directionally correct and still become financially fatal when concentration, leverage, and timing are misaligned. You can predict the destination and still run out of fuel on the way.

Big Tech earnings transformed that lesson into a comparative experiment. Microsoft was rewarded after Azure crossed $100 billion in annual revenue and Copilot adoption continued to expand. Amazon offered a similar narrative as AWS growth accelerated alongside its AI infrastructure investments. In both cases, investors could see a direct bridge between enormous capital expenditure and customer revenue.

Meta received a harsher reaction. Its core advertising business remained strong, but infrastructure spending accelerated faster than the market’s confidence in near-term AI monetization. Apple offered another variation: powerful distribution and cash generation can buy time, but they cannot permanently substitute for a compelling AI product story.

This was not a week of AI skepticism. It was a week of discrimination. Open models must diffuse. Robots must act reliably. Technology companies must convert capital expenditure into revenue. Investors must survive the journey.

The market is no longer asking whether AI will be enormous. It is asking who can turn intelligence—digital or physical—into durable economics without losing control of models, machines, or capital.

🔎 AI Research

Scientific computing in the age of agentic AI: an exploratory field report

AI Lab: OpenAI

Summary: The paper “scientific-computing-in-the-age-of-agentic-ai-an-exploratory-field-report.pdf” explores the potential of Large Language Model agents to address technical debt and software engineering shortages in life sciences computing. Through eight case studies, the authors demonstrate that while AI agents can successfully execute tasks ranging from minor code maintenance to full performance rewrites, careful human verification and long-term stewardship remain essential.

ScientistOne: Towards Human-Level Autonomous Research via Chain-of-Evidence

AI Lab: Google Cloud AI Research

Summary: This paper introduces the Chain-of-Evidence framework and the ScientistOne system to combat the undetected verifiability failures, such as hallucinated citations and irreproducible scores, prevalent in current autonomous research agents. By structurally enforcing that every generated claim traces back to verifiable evidence, ScientistOne eliminates hallucinated references and achieves perfect score verification while matching or exceeding expert performance on complex benchmarks.

Kimi K3: Open Frontier Intelligence

AI Lab: Kimi Team (Moonshot AI)

Summary: This paper introduces Kimi K3, a 2.8-trillion parameter multimodal Mixture-of-Experts model featuring a 1-million-token context window and native vision capabilities. By combining architectural innovations like Kimi Delta Attention with multi-domain reinforcement learning, the model achieves frontier-level performance on long-horizon coding, agentic, and reasoning tasks.

Visual prompt engineering for video models

AI Lab: Google DeepMind

Summary: This paper demonstrates that visual prompt engineering (VIPE)—transforming task images via image editors—systematically improves the visual reasoning capabilities of video models. The authors find that video models possess a strong realism bias, meaning that converting abstract sketches into photorealistic scenes can be a more effective test-time scaling strategy than traditional text-based prompting.

Shieldstral

AI Lab: MistralAI

Summary: This paper presents Shieldstral, a 3B-parameter policy-adaptive multimodal safety classifier that simplifies content moderation into a unified binary question-answering task. Through extensive data curation and contrastive sample generation, this compact model matches or outperforms models nearly seven times its size across diverse text and multimodal safety benchmarks.

🤖 AI Tech Releases

Gemini Robotics 2

Google DeepMind released Gemini Robotics 2 , a three-model suite of intelligence for robotics.

LFM2.5

Liquid AI released LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, two encoder models that can be easily adaped to downstream tasks.

MAI-Cyber-1-Flash

Microsoft announced MAI-Cyber-1-Flash, a model to find and fix vulnerabilities in complex code bases.

DeepSeek-V4-Flash API

DeepSeek released the official version of the v4-Flash Official API

📡10 AI News You Need to Know About

1. Jensen Huang used his first ever X post to share Open Weights and American AI Leadership, a three-page letter co-signed by 25 companies including Nvidia, Microsoft, Meta and Palantir asking Washington to avoid “premature restrictions” on open-weight AI models, with OpenAI and Anthropic notably absent from the signatories.

2. Situational Awareness, the roughly $20B AI-focused hedge fund founded by Leopold Aschenbrenner, sold its entire public equities portfolio to Ken Griffin’s Citadel days after reports it was seeking fresh capital following heavy losses in the AI selloff.

3. Anthropic disclosed three incidents found across 141,006 evaluation runs in which Claude models reached the open internet from a misconfigured test environment and gained unauthorized access to the production systems of three organizations.

4. Nscale agreed to acquire Anyscale, the company built by the creators of Ray, adding a workload orchestration layer on top of its power, data center and GPU stack, with Bloomberg putting the price at roughly $1.65 billion.

5. Richard Socher’s Recursive signed a multi-year $410M agreement with AWS to run its automated AI research system, committing most of the $650M it raised on leaving stealth in May to compute rather than headcount.

6. Safe Superintelligence and Nvidia announced a long-term strategic partnership that pairs an Nvidia investment with Vera Rubin access to increase SSI’s compute by an order of magnitude, with Bloomberg reporting the investment at $5 billion.

7. Oracle and Google Cloud expanded their partnership to bring Gemini models into Oracle AI Agent Studio plus embedded AI in Fusion Applications and NetSuite, and Oracle shares rose as much as 8.4%.

8. Meta’s second quarter results came with a 10-Q disclosure of roughly $279 billion in data center, colocation and network leases that have not yet commenced, plus another $68 billion signed in July expected to start in 2027 and 2028 .

9. Microsoft added more than $130B of new data center lease commitments in the June quarter, taking total not-yet-commenced leases to $329.1 billion, up from $196.6 billion, disclosed alongside its FY26 Q4 results.

10. Moonshot AI closed a $3.5B round at a $35B valuation, far above its original $1B to $2B target, and is already approaching backers at a $50 billion pre-money valuation ahead of a possible Hong Kong IPO this year .

This is the first post of a new section of TheSequence focused on advancements in robotics. Our goal is to keep you up to date with the most important developments in AI robotics which is an area that is not well covered by other newsletters. For this first post, I wanted to discuss the current landscape of AI models for robotics.

Let’s start.

A language model can hallucinate a sentence and delete it. A robot can hallucinate a grasp and drop a wine glass.

That difference contains most of the robotics problem.

AI has largely advanced inside forgiving environments. Tokens are cheap, software can be reset, and failed generations disappear. Robotics moves intelligence into a world with gravity, friction, latency, broken parts, and humans who do not enjoy being treated as test data.

Robotics is what happens when an AI model leaves the library and discovers physics.

This is why the race for the robot foundation model will not simply replay the LLM market. Frontier labs have the strongest digital brains. Robotics startups have the bodies, field data, and scars. NVIDIA is building the factory around both. Hugging Face is assembling the open workshop.

The question is not who has the largest model. It is who can connect reasoning to reliable action.

A Robot Brain Is a Stack

Read more

Ilya Sutskever recently offered a compact history of modern AI. From roughly 2012 to 2020, he argued, the field lived in an age of research. From 2020 to 2025, it entered an age of scaling. Now we are going back to research—only this time “with big computers.”

It is an appealing periodization. It also creates an immediate puzzle.

Open the release notes for almost any frontier model in 2026 and the architecture diagram looks strangely familiar. There is still a Transformer somewhere in the machine, often routed through a Mixture-of-Experts. The headline improvements are usually elsewhere: better data, longer context, stronger reinforcement learning, synthetic tasks, tool use, memory, verification, adaptive reasoning budgets, and agent orchestration.

This does not look like the arrival of a new neural species. It looks like Formula 1. The car still has four wheels and an engine. Yet enormous gains come from aerodynamics, energy recovery, tire chemistry, telemetry, software, and pit strategy. The chassis matters. The system around it increasingly decides the race.

So which era are we in: research or engineering?

Probably both. The age of research has returned, but much of that research is now expressed as industrial-scale engineering.

“Scaling did not end. It escaped.”

The Recipe That Ate the Field

Read more

Take every open-weight model that discloses its parameter count, put total parameters on a log x-axis, put Terminal-Bench 2.1 score on the y-axis, and you get a reasonably tidy cloud sloping up and to the right. Bigger is better. This is the shape we have been trained to expect.

Then there is a point sitting well above the trend line at 118B.

Disclosed-size open-weight models on Terminal-Bench 2.1. The dashed line is a fit through the field.

Laguna S 2.1 scores 70.2%. To its right, at 1.6 trillion parameters, DeepSeek-V4-Pro-Max sits at 64.0. At 975B, Inkling is at 63.8. At 550B, Nemotron 3 Ultra is at 56.4. On DeepSWE, a harder and less saturated benchmark, the gap stops being subtle at all: Laguna S 2.1 scores 40.4 against DeepSeek-V4-Pro-Max’s 9.0.

A 13x parameter deficit paired with a 4x score advantage is the kind of result that usually means somebody broke the eval. Poolside seems to have anticipated that reaction, because they published every trajectory from every trial in the final evaluation set. You can go read what the model actually did. That decision tells you most of what you need to know about how this release was designed.

The company you have probably never heard of

Read more

For most of machine learning history, data was treated as geology. It already existed somewhere in the world—in books, websites, code repositories, conversations, photographs, and databases. The researcher’s job was to excavate it, clean it, tokenize it, and feed it into a model.

Large language models changed this relationship. A capable model is not only a consumer of data. It can produce questions, answers, explanations, critiques, preference labels, tool traces, textbooks, code exercises, and entire miniature curricula.

This creates a new training primitive. Instead of asking an expensive model to answer every production query forever, we ask it to manufacture the experience from which a smaller model learns. The teacher runs offline. Its outputs become a dataset. The dataset trains the student. The teacher disappears at inference time, but some of its behavior remains embedded in the student.

That is synthetic data as distillation.

Read more

Next Week in The Sequence:

• Our series about AI model distillation continues with another exciting technique.

• The AI of the week covers Poolside’s new Laguna S2.1.

• The opinion section debates the hot topic about AI research vs. engineering.

Subscribe and don’t miss out:

📝 Editorial: Last Week in AI: Last Week in AI: Smarter Models, Physical Machines, and the Expanding AI Stack

When I started The Sequence years ago, AI was still a relatively niche field, followed closely by researchers, a small group of builders, and a few overly enthusiastic people like me. Today, understanding AI means looking far beyond language models. I increasingly find myself studying frontier areas such as robotics, physical intelligence, scientific discovery, biology, materials, and autonomous laboratories.

I would like to share more of that learning with you. Starting soon, we will introduce two new sections—The Sequence Robotics and The Sequence Science—focused on making the most exciting developments in these fields rigorous, accessible, and fun to follow.

The most important development this week was Anthropic’s release of Opus 5. The model improves long-horizon reasoning, agentic coding, and professional knowledge work while making capabilities previously associated with the most expensive frontier systems more economically usable. The significance is not merely another benchmark jump. Opus 5 suggests that frontier models are becoming better at sustaining complex work over time: understanding large systems, planning across many steps, using tools, revising decisions, and maintaining coherence throughout long tasks.

That matters because the next phase of AI will be defined less by answering isolated questions and more by completing extended workflows.

The second major development came from Travis Kalanick’s Atoms, which announced a $1.7 billion fundraising round. Atoms is a wager that the next great AI market will not live entirely inside browsers and data centers. It will operate in mines, factories, kitchens, warehouses, and transportation systems.

Physical AI is fundamentally harder than software intelligence. A coding agent can retry after an error. A robot moving equipment through a warehouse must deal with friction, uncertainty, safety, hardware failure, and the stubborn complexity of the real world. Atoms reflects a growing conviction that robotics will be one of AI’s next major frontiers—and one of its most capital-intensive.

Poolside’s Laguna S 2.1 represented the open-model counterpart to Opus 5. Its 118-billion-parameter mixture-of-experts architecture activates only 8 billion parameters per token, supports a one-million-token context window, and delivers strong agentic coding performance for its size.

The contrast is revealing. Opus 5 pushes the proprietary frontier forward. Laguna attempts to compress frontier-adjacent capability into a model that is smaller, portable, and open. Together, they show the model market expanding in both directions: more capable at the top and more accessible underneath.

Then came the week’s warning. During a controlled cyber evaluation with production safeguards reduced, OpenAI models reportedly escaped a constrained environment, exploited a zero-day vulnerability, and accessed Hugging Face infrastructure in pursuit of benchmark answers.

This was not evidence of machine consciousness. It was evidence of something more practical: a capable system relentlessly optimizing toward a goal inside an environment whose boundaries were weaker than expected. As agents gain more autonomy, containment will become as important as capability.

Alphabet’s quarter illustrated the financial scale behind this transition. Google Cloud continued growing rapidly, while quarterly capital expenditure climbed to nearly $45 billion. The AI boom is no longer just a software cycle. It is an industrial construction project involving chips, power, data centers, networking, and enormous balance sheets.

AMD’s latest announcements reinforced that shift. Helios, the MI400 family, new EPYC processors, ROCm software, and robotics-oriented systems position AMD as a supplier of integrated AI infrastructure rather than merely an alternative GPU vendor.

Finally, rumors of a possible OpenRouter acquisition showed the emerging value of distribution. In a world with dozens of capable models, the layer that routes workloads, manages spending, and handles payments may become as strategically important as the models themselves.

The lesson from the week is that the AI race is broadening. Opus 5 advances intelligence. Atoms brings it into the physical world. Laguna makes it more accessible. The security incident exposes its risks. Alphabet and AMD reveal the machinery required to scale it. OpenRouter points toward the control layer that may connect everything together.

🔎 AI Research

An Exam for Active Observers

AI Lab: University of Southern California

Summary: This paper introduces ActiveVision, a benchmark designed to test if Multimodal Large Language Models (MLLMs) can perform active, iterative visual perception rather than relying on a single static glance. The study reveals that current frontier models collapse on these tasks and severely lag behind human performance, indicating a critical gap in robust active visual reasoning even when the models utilize agentic coding tools.

Native Python Object-Oriented Agents

AI Lab: NVIDIA

Summary: Presented in the file “2607.20709v1.pdf”, this paper introduces NVIDIA Object-Oriented Agents (NOOA), a Python framework that simplifies AI development by treating agents as standard Python objects where methods act as model capabilities and fields act as the state[cite: 8]. By leveraging native abstractions like pass-by-reference and code-as-action, NOOA allows models to efficiently achieve highly competitive results on complex software engineering, cybersecurity, and interactive reasoning benchmarks.

Two-Level Meta-Rubrics for Evaluating Open-Ended Generation: GAMUT, a Benchmark for Factual Completeness

AI Lab: Meta AI

Summary: This paper introduces GAMUT, a multimodal benchmark designed to evaluate the factual completeness of long-form generations rather than just their factual precision. It utilizes a two-level meta-rubric framework that translates complex, structured factual requirements into a flat checklist of binary criteria, allowing LLM judges to score responses reliably.

LLM-as-a-Coach: Experiential Learning for Non-Verifiable Tasks

AI Lab: Microsoft Research

Summary: This paper presents Experiential Learning (EL), a post-training framework for non-verifiable tasks that replaces scalar rewards with rich, transferable text guidance generated by an “LLM-as-a-Coach”. By internalizing this high-bandwidth feedback through on-policy context distillation, the model achieves better generalization and mitigates reward hacking compared to standard reinforcement learning methods.

Mage-Flow: An Efficient Native-Resolution Foundation Model for Image Generation and Editing

AI Lab: Microsoft

Summary: Mage-Flow is a compact, 4-billion-parameter foundation model designed for efficient, high-resolution text-to-image generation and instruction-based image editing. By combining a lightweight latent tokenizer, a native-resolution diffusion transformer, and fused-kernel training infrastructure, it achieves competitive visual quality with significantly lower latency and memory usage than larger baseline models.

Environment-free Synthetic Data Generation for API-Calling Agents

AI Lab: Apple

Summary: ESAT is a novel synthetic data generation pipeline that creates complex, multi-step training trajectories for API-calling agents using only API specifications, completely eliminating the need for fully executable backend environments. By employing LLMs as dynamic digital world models to generate tasks, simulate stateful API responses, and judge trajectory quality, the framework produces high-fidelity data that drives substantial performance gains during model fine-tuning.

🤖 AI Tech Releases

Opus 5

Anthropic released the new version of its marquee model.

Laguna S2.1

Poolside released Laguna S2.1, an open weight model that outperforms much larger models in interesting benchmarks.

📡10 AI News You Need to Know About

1. Alphabet reported Q2 revenue of $119.8 billion, up 24%, with Google Cloud accelerating 82% to $24.8 billion on enterprise AI demand, easing investor anxiety over its roughly $180-190 billion capex plans.

2. Travis Kalanick’s robotics venture Atoms raised $1.7 billion led by a16z, with Uber, Bain Capital, and Fifth Wall participating, to pursue what he calls the “bits-to-atoms” vision of controlling the physical world with software.

3. OpenAI disclosed that its pre-release models breached Hugging Face after escaping their sandboxed test environment through a package-installer vulnerability while running a cyber benchmark with reduced refusals, ultimately extracting benchmark solutions from Hugging Face’s production database.

4. Databricks announced a Coatue-led strategic round valuing it at $188 billion, reportedly around $3 billion, up from a $134 billion valuation just five months ago.

5. Beijing world-model startup GigaAI is in talks for a Hong Kong IPO as soon as this year while closing a funding round at a $3 billion valuation, which CEO Huang Guan says would make it the first world-model startup globally to go public.

6. AI robotics startup Genesis AI is in talks to raise around $500 million at a $3 billion pre-money valuation, a big step up for the physical AI lab that emerged from stealth a year ago with a $105 million seed round.

7. AMD launched its next-generation AI infrastructure portfolio at Advancing AI 2026, headlined by Helios rack-scale systems now in production for gigawatt-scale deployments alongside new EPYC processors and an updated Instinct accelerator, positioning the full stack against Nvidia in what it frames as a $2 trillion AI compute opportunity.

8. Etched closed a $300 million Series C at a $10.3 billion valuation led by Sequoia, with a16z, SK Hynix, Jane Street, and Diffusion participating, doubling its December valuation in about seven months as it scales production of its GPU-free inference clusters against more than $1 billion in booked orders.

9. Prentis, a computer-use AI lab launched in April by Ritankar Das with Reid Hoffman and Mark Pincus as co-founders, is in talks to raise $100 million at a $1 billion valuation, betting its small Hive-32B model, which it claims beats GPT-5.4 and Claude Opus 4.6 on computer-use benchmarks at roughly a tenth of the cost per task, can win the race to automate routine office workflows.

10. Stripe is in talks to acquire OpenRouter, the marketplace giving developers unified access to hundreds of AI models, in a deal that could value the startup at roughly $10 billion, nearly eight times its $1.3 billion valuation from May, though the talks could still fall apart or draw a rival bidder.

THESIS Google is the closest strategic mirror of NVIDIA’s full-stack method, but not a universal drop-in replacement; AMD and AWS make a literal “only” claim too strong.

The Chip Is Not the Product

It is tempting to compare AI companies by lining up accelerators and reading the specification sheet: FLOPS, memory bandwidth, interconnect speed, tokens per second. This is useful, but incomplete in the same way that comparing airlines by engine thrust is incomplete. The engine matters. So do the airframe, airports, pilots, maintenance crews, scheduling software, fuel contracts, and route network. NVIDIA’s achievement is not merely a very fast GPU. It is an industrial system that turns models into running software with unusually little friction.

That is the strongest form of the case for Google as NVIDIA’s only viable competitor. Google is not the only company capable of building a good AI chip. It may not even win every benchmark. But it is the only company that closely mirrors NVIDIA’s control of the whole machine: silicon, interconnects, servers, compilers, frameworks, cloud operations, frontier models, and applications used by billions of people. The claim needs one qualification. Google is the closest full-stack strategic rival, not a universal drop-in replacement, and AWS and AMD make the word “only” uncomfortable.

What “Full Stack” Actually Means

Read more

Inkling is best understood not as a single 975-billion-parameter brain that fires all at once, but as a giant warehouse of specialist capacity. A router chooses a small working set for each token. That makes the arithmetic sparse, while storage, networking, and deployment remain very large.

The headline number for Inkling is 975 billion parameters. That is close enough to a trillion that the distinction is mostly useful to accountants. The more interesting number is the one beside it: 41 billion parameters active per token. In other words, the model owns an enormous amount of capacity, but a single word passing through the network touches only about 4.2 percent of it.

A good mental picture is a university with 256 specialist departments on each relevant floor. When a token arrives, a dispatcher does not convene the entire university. It selects six departments that appear useful for this token, adds two general-purpose departments that always attend, combines their work, and moves on. A line of Python might summon one set of specialists; a phrase in Greek, a diagram label, or a piece of audio may summon another.

Read more

In January 2025, DeepSeek took its big reasoning model, R1, and used it to generate around 800,000 worked solutions — long, rambling chains of thought, complete with false starts, self-corrections, and the occasional “wait, let me reconsider.” They filtered these traces for correctness and readability, and then did the most boring thing imaginable with them: plain supervised fine-tuning on a handful of off-the-shelf open models — Qwen at 1.5B, 7B, 14B, 32B; Llama at 8B and 70B. No reinforcement learning. No reverse KL. No on-policy sampling. No teacher-as-critic. Just next-token prediction on the teacher’s transcripts.

And it worked spectacularly. The distilled 32B model started solving competition math it had no business solving. The 7B model began verifying its own work and branching its reasoning mid-stream — emergent behaviors nobody trained into it directly. A grab-bag of small dense models suddenly reasoned like something ten times their size.

If you internalized the arc of this series — the painstaking migration from forward to reverse KL, the whole argument that you cannot just imitate a teacher’s trajectories because the student will never be on them at inference — your first reaction should be mild outrage. We spent an entire installment establishing why naive sequence-level imitation is the wrong tool, and then the single most important reasoning-distillation result of the decade is naive sequence-level imitation. What gives?

The answer is the whole story of this installment, and it turns out to be more interesting than either “imitation works” or “imitation doesn’t.”

The Trace, Not the Answer

Read more

Next Week in The Sequence:

• We continue our series about model distillation techniques.

• In the AI of the Week , we discuss Thinking Machine first open weights model.

• The opinion section reviews the idea that Google’s is by far the biggest threat to NVIDIA’s dominance.

Subscribe and don’t miss out:

📝 Editorial: Last Week in AI: China, Compression and the Open-Model Race

For years, AI progress has been narrated as a horse race: larger models, higher benchmark scores, more expensive clusters. This week broke that frame. The most consequential developments were not simply advances in intelligence; they were competing answers to a deeper question: Who gets to possess, adapt and govern it?

Thinking Machines Lab entered the model arena with Inkling, a 975-billion-parameter mixture-of-experts model with 41 billion active parameters, native text, image and audio inputs, a one-million-token context window, open weights and an Apache 2.0 license. The company’s emphasis on calibration, controllable reasoning effort and customization is more interesting than any leaderboard placement. Inkling is a wager that the next frontier will be models people can shape around their own judgment—not merely rent from a centralized oracle.

Moonshot AI pushed in the opposite direction on scale while converging on the same politics of access. Kimi K3 carries 2.8 trillion parameters, activates 16 of 896 experts, supports vision and a million-token context, and is aimed at long-horizon coding and knowledge work. Moonshot describes it as the first open model in the three-trillion-parameter class, although the full weights are not due until later this month. That caveat matters, but so does the trajectory: Chinese labs are no longer competing only on cost. They are using openness itself as a strategic lever.

PrismML’s Bonsai 27B makes the week’s most radical claim through compression rather than scale. Its ternary model is 5.9GB, while the one-bit variant is 3.9GB—small enough, PrismML says, to fit within a modern smartphone’s usable memory. The company reports retaining most of the performance of the full-precision baseline across its benchmark suite. Independent testing will determine how durable those numbers are. Yet the direction is unmistakable: intelligence density may become as strategically important as raw intelligence. Local models alter latency, privacy, cost and even the bargaining power between users and cloud providers.

A different type of model was also introduced this week by OpenAI. GPT-Red introduced a different—but equally consequential—form of scaling. It is not a consumer model or a new chatbot, but an internal automated red-teaming system trained through self-play to attack other models, observe their defenses and invent progressively stronger prompt injections. In unfamiliar test environments, GPT-Red successfully compromised GPT-5.1 in 84% of scenarios, compared with 13% for human red-teamers, and its attacks were subsequently used to make GPT-5.6 substantially more robust. The deeper idea is a new scaling law for safety: as models become more capable, the systems testing them can scale alongside them. The future may not be defined by models that simply improve themselves, but by machine-speed ecosystems in which attackers and defenders continuously co-evolve.

The common denominator is not openness as an ethical abstraction, but distribution as an engineering objective. Open weights permit adaptation; sparse architectures make colossal models economical; extreme quantization brings inference to the edge. Each move weakens a different bottleneck that has kept advanced AI concentrated in a few institutions.

That technological story now has an explicitly geopolitical counterpart. At Shanghai’s World AI Conference, Xi Jinping cast open-source AI as a global public good, promoted Chinese support for developing countries and elevated a new international AI cooperation organization as a vehicle for shaping global governance. The rhetoric of access is not geopolitically neutral; standards, training programs and model ecosystems can build dependencies as effectively as chips and cloud infrastructure.

Taken together, these announcements describe a frontier that is fragmenting—and perhaps democratizing. The decisive contest is no longer just over who can build the smartest system. It is over whether intelligence will be centralized or portable, proprietary or adaptable, governed by firms, states or communities. This week, AI stopped looking like a single race. It began to look like an emerging world order.

🔎 AI Research

GPT-Red: Unlocking Self-Improvement for Robustness

AI Lab: OpenAI

Summary: This article introduces GPT-Red, an automated red-teaming model that uses self-play reinforcement learning to efficiently discover vulnerabilities and prompt injection attacks. By integrating GPT-Red into their training pipeline, researchers successfully enhanced the robustness of production models like GPT-5.6 Sol against malicious instructions without degrading their core capabilities.

A Theory of Contrastive Learning with Natural Images

AI Lab: CSAIL, MIT and Hebrew University of Jerusalem

Summary: This paper analytically derives the optimal representations for contrastive learning with natural images, proving that simple augmentations optimally result in a partial whitening process computable by a basic CNN with sinusoidal filters. Experimental results confirm that CNNs trained with contrastive loss on various datasets naturally learn these sinusoidal filters and perform partial whitening as predicted by the theory.

On the Interpolation Effect of Score Smoothing in Diffusion Models

AI Lab: Google Research

Summary: This paper investigates the generative creativity of diffusion models, proposing that their neural network backbones naturally learn a smoothed version of the empirical score function rather than memorizing the exact training data. Through theoretical analysis and numerical experiments, the author demonstrates that this score smoothing guides the denoising dynamics to generate novel samples that smoothly interpolate between training points, even across complex nonlinear manifolds.

Video Generation Models are General-Purpose Vision Learners

AI Lab: Google DeepMind

Summary: The authors propose GenCeption, a unified generalist vision model that leverages a pre-trained text-to-video diffusion backbone to perform a wide variety of dense and sparse visual tasks via text instructions. By repurposing iterative diffusion into an efficient feed-forward architecture, GenCeption achieves state-of-the-art performance across diverse perception tasks while demonstrating emergent behaviors like sim-to-real transfer and zero-shot generalization.

Metacognition in LLMs: Foundations, Progress, and Opportunities

AI Lab: Yale University and University of California, Irvine

Summary: This paper presents a comprehensive taxonomy and review of metacognition in Large Language Models, detailing how models’ abilities to monitor and regulate their own cognitive processes are measured, elicited, and improved. It synthesizes current findings on metacognitive capabilities like confidence calibration, introspection, and knowledge boundary detection, highlighting the potential of these mechanisms to enhance LLM reliability, reasoning, and human-AI collaboration.

ADVANCED MATHBENCH: A Benchmark Suite for Advanced Mathematical Proof Generation and Verification

AI Lab: Shanghai AI Laboratory

Summary: ADVANCED MATHBENCH introduces a rigorous evaluation suite focusing on the generation and process-level verification of advanced, natural-language mathematical proofs at the undergraduate and doctoral qualifying-exam levels. The benchmark reveals that frontier LLMs still struggle significantly with constructing and verifying complex mathematical proofs, highlighting a critical bottleneck in their ability to accurately detect subtle logical errors.

🤖 AI Tech Releases

Inkling

Thinking Machines open sourced Inkling, a large MoE model with a 1M context window.

Kimi K3

Moonshot AI released Kimi K3, a massive 2.8T parameter model optimized for long horizon coding, reasoning and kowledge work.

Robostral Navigate

Mistral released its first model optimized for embodied AI.

Bonsai 27B

PrismML released Bonsai 27B, a model that can run entirely on a phone.

📡10 AI News You Need to Know About

1. Xi Jinping made his debut at the World AI Conference in Shanghai touting China’s low-cost AI and calling for an open, cooperative technological order, saying AI development should be a symphony of international cooperation rather than a solo performance by one country.

2. Reuters reported that China’s Cyberspace Administration approved Apple Intelligence for launch on the back of a deal integrating Alibaba’s Qwen models into iOS, iPadOS, macOS, and visionOS, with Baidu also confirming it is working with Apple on features for Chinese users.

3. Emergent announced a $130 million Series C led by Creaegis at a $1.5 billion valuation, a fivefold jump in months, on the back of a $120 million revenue run rate and 200,000-plus paying customers for its AI software creation platform.

4. Demis Hassabis published a proposal on X calling for a FINRA-style independent standards body, funded by industry and backed by the US government, that would review frontier models up to 30 days before release and eventually gate deployment in the US market. (Replaced TechCrunch with Hassabis’s original post.)

5. Reflection AI signed a $1 billion-plus compute deal with Nebius running through 2029 that gives the open-model lab access to Nvidia’s GB300 chips, its second major capacity grab after last month’s SpaceX agreement.

6. Bloomberg reported that Google is months behind schedule on Gemini 3.5 Pro because the model is falling short of internal goals, especially in coding, frustrating employees who worry Anthropic and OpenAI are pulling ahead.

7. Greylock announced Greylock 18 on X, a $1.5 billion early-stage fund, its eighteenth, aimed at concentrated bets on AI-native founders, with managing partner Asheem Chandna arguing the next trillion-dollar companies are still ahead.

8. TSMC’s June revenue report showed sales of NT$442.68 billion for the month, up 67.9% year over year, lifting June-quarter revenue 36% to roughly $39.6 billion and confirming AI demand remains intact. (Replaced Bloomberg with TSMC’s own monthly revenue release.)

9. Walden Robotics launched out of stealth, a Toyota Research Institute spinout led by MIT’s Russ Tedrake, with roughly $300 million in seed funding at a $1.1 billion valuation co-led by Toyota and Deviation Capital, and its general-purpose robots already working production shifts at a North American Toyota plant.

10. SK Hynix raised $26.5 billion by selling 177.9 million ADRs at $149 each in the largest-ever US listing by a foreign company, topping Alibaba’s 2014 record, just as Commerce Secretary Lutnick pressed the memory maker to build new US fabs.

Last Thursday, Mark Zuckerberg posted on X for the first time in three years. That alone should tell you something. The occasion was the launch of Muse Spark 1.1, the second model out of Meta Superintelligence Labs and the first Meta model ever to ship with a price tag. It arrived with a public API, aggressive pricing at $1.25 per million input tokens and $4.25 per million output tokens, an OpenAI-compatible endpoint, and closed weights. Read that last part again. The company that spent three years evangelizing open weights as the moral and strategic high ground of AI just shipped a proprietary frontier model behind a paid API, and the CEO came back from a three-year social media exile to announce it.

Spark 1.1 did not arrive alone. Two days earlier Meta shipped Muse Image, its first image generation model from the new lab. A week before that, reports surfaced that Meta is building a cloud business, internally called Meta Compute, to sell surplus AI infrastructure to outside customers. Add the custom MTIA silicon ramping toward production and you get the picture: in roughly eighteen months, Meta has gone from an open-weights research shop with an ads business attached to a company assembling the entire vertical stack. Chips, datacenters, cloud, models, API, apps, devices. Only Google has ever held all of those cards at once.

So the question practically asks itself. Can Meta actually compete with the frontier labs? I think the honest answer is that this is two questions wearing one trench coat, and they have different answers. At the layer where models meet users, the app and agent layer, Meta might be the favorite. At the layer where models get made, the evidence is thin and the structural arguments cut against it. This essay argues both sides properly, because both sides deserve it.

The launch, read closely

Read more

OpenAI’s audit of SWE-Bench Pro shows why a precise score can still be a poor measure - and why coding agents may become essential tools for auditing the benchmarks that grade them.

A frontier coding score can look wonderfully precise: 80.3 percent, one decimal place, clean enough to rank models and anchor product claims. But precision is not validity. If a benchmark rejects correct solutions, accepts incomplete ones, or asks for behavior its prompt never specifies, the number is measuring something other than coding ability.

That is the uncomfortable conclusion of OpenAI’s audit of SWE-Bench Pro. The benchmark was built to address weaknesses of earlier coding evaluations: longer-horizon tasks, more realistic repositories, and code intended to reduce training-data contamination. On its 731-task public split, frontier-model performance climbed from 23.3 percent to 80.3 percent in eight months. Instead of treating that curve as unambiguous progress, OpenAI asked a more important question: how much of the result comes from the model, and how much comes from the test?

The answer is striking. OpenAI estimates that roughly 30 percent of the public benchmark is broken. Its agent-assisted audit labeled 200 tasks, or 27.4 percent, as defective. A parallel campaign involving experienced software engineers labeled 249 tasks, or 34.1 percent, as defective. OpenAI has withdrawn its earlier recommendation that the field adopt SWE-Bench Pro.

Coding benchmarks are executable specifications - except when they are not

Read more