For this week in AI’s essay, I would like to discuss Thinking Machines’ work on interactive models which takes multi-modality to a new level. I’ve been diving as much as possible on their ideas and wanted to share some thoughts. The work is early but so impressive. Check it out to get started:

For the last few years, the default mental model for large language models has been embarrassingly simple: concatenate tokens, predict the next token, repeat. The human writes a message, the model replies, the human writes again. This works surprisingly well for many tasks because text is forgiving. Text can wait. It can be buffered, edited, compressed, and serialized into one neat causal stream.

But collaboration is not text. Collaboration is temporal.

Read more

💡 AI Concept of the Day: What are Text Diffusion Models

If you look at the architecture of the modern AI boom, it is heavily bifurcated by modality. In the visual domain, we are entirely ruled by diffusion models. From Midjourney to Stable Diffusion to OpenAI’s Sora, the paradigm of starting with pure noise and iteratively denoising it into a high-fidelity image or video has proven to be unreasonably effective.

But in the realm of text, diffusion has historically been an afterthought. Large Language Models (LLMs) like GPT-4, Claude, and LLaMA are staunchly autoregressive (AR). They are sequence predictors. They look at the context, predict the next token, append it to the context, and repeat. It is a strictly left-to-right, causal process.

For years, the consensus was simple: autoregression is just the native physics of language. But this sequential paradigm has glaring pathologies. Because AR models generate blindly from left to right, they cannot easily engage in global planning. If they make a slight logical error early in a sequence, that error is committed to the context window permanently, leading to cascading failures—a phenomenon often critiqued as “generation drift.” Furthermore, AR models suffer from the “reversal curse”; they can easily recite a poem forward, but if you ask them to recite it backward, their causal attention mechanisms break down entirely.

Read more

Next Week in The Sequence:

1. We continue our series about trasnformer alternatives.

2. In our opinion section, we discuss the thesis that “every agent needs a computer”

3. In the AI of the Week section, we dive into Thinking Machines’ interactive models which can think and listen and same time.

Subscribe and don’t miss out:

📝 Editorial: Last Week in AI: IPOs, Interactive Models, and Recursive Dreams

This week in AI felt less like a product cycle and more like a philosophical provocation wrapped in a market event, a demo, and a few delightfully ambitious lab announcements. The common thread was not bigger models, larger context windows, or yet another benchmark victory. It was agency. Who gets to shape intelligence? Who gets to improve it? And, slightly more ominously, what happens when the tools begin improving the tools?

Start with Cerebras. The IPO was not just a financing milestone for an AI chip company; it was a reminder that the AI race is still brutally physical. Behind every magical chatbot sits an industrial stack of silicon, power, networking, cooling, capital markets, and geopolitical anxiety. Cerebras has always been the wonderfully weird character in the AI hardware opera: instead of making chips modestly larger, it went all-in on the wafer-scale computer, basically asking, “What if the chip were the data center?” That sounds like something invented after too much espresso at a semiconductor conference, but it captures an important truth. If intelligence is becoming a commodity, compute remains the refinery. Public markets are now voting on which refineries matter.

Then came Thinking Machines with its interaction models, and the shift in mood was immediate. While much of the industry keeps celebrating autonomous agents that disappear into the cloud and return with a spreadsheet, Thinking Machines is betting on the opposite: AI that stays present. Not “prompt, wait, receive,” but continuous, multimodal collaboration. The model listens, watches, interrupts, yields, and reacts in real time. In other words, it treats interaction not as a UI layer bolted onto intelligence, but as part of intelligence itself. This is a subtle but profound idea. Human cognition is social before it is computational. We think through gestures, pauses, corrections, shared context, and awkward silences. Making AI more useful may require making it less like an oracle and more like a very fast, very attentive collaborator who knows when not to talk.

The most intellectually spicy developments, however, came from the new “AI scientist” movement. Recursive emerged with the grand ambition of building systems that improve themselves through automated experimentation. Adaption’s AutoScientist pointed in a similarly important direction, automating the loop behind training, alignment, and model adaptation. The dream here is intoxicating: research no longer as a linear human process, but as a compounding machine of hypotheses, experiments, evaluations, and refinements.

That dream also deserves a raised eyebrow. Recursive self-improvement has long lived somewhere between computer science, science fiction, and existential risk seminar. Turning it into a venture-backed product category is both thrilling and slightly unhinged — which, frankly, is how many important technologies begin. The question is not whether AI can help with research. It already does. The question is whether we can build systems that improve capabilities without eroding grounding, safety, diversity, or human judgment.

Another signal came from China, where Junyang Lin, the former lead researcher behind Alibaba’s Qwen models, is reportedly raising several hundred million dollars for a new AI lab at a valuation that could land around $2 billion. That number is almost rude for a brand-new lab — especially in China, where AI startup valuations tend to be more earthbound than in the U.S. — but Lin has the rare credential that matters in frontier AI: he has already helped turn a model family into an ecosystem. His departure also exposes a delicious tension inside Chinese AI: as Alibaba shifts more energy toward proprietary models and monetization, the open-source talent orbit may be ready to spin out into independent labs. But this is not simply a “raise money, buy GPUs, train model, repeat” story. New Chinese AI labs must navigate U.S. chip export controls, constrained compute access, and the tricky problem of building research agendas that do not merely overlap with giants like Alibaba and ByteDance. In that sense, Lin’s new lab is less another shiny founder myth than a test case for whether China’s open-model momentum can become a venture-scale company.

The machine is no longer just answering. It is listening, adapting, experimenting — and, perhaps, beginning to ask what it should become next.

🔎 AI Research

MemEye: A Visual-Centric Evaluation Framework for Multimodal Agent Memory

AI Lab: Rutgers, Notre Dame, Princeton, UMN, SAMD

Summary: MemEye is a novel evaluation framework and benchmark designed to assess whether multimodal AI agents can effectively preserve and reason over fine-grained visual evidence in long-term interactions. By evaluating 13 memory methods, the authors reveal a critical trade-off where text-based memory loses visual details while native image memory struggles with tracking temporal state changes.

EVOLVEMEM: Self-Evolving Memory Architecture via AutoResearch for LLM Agents

AI Lab: UNC-Chapel Hill, UC Berkeley, UCSC

Summary: EVOLVEMEM introduces a self-evolving memory architecture that uses an LLM-powered diagnosis module to autonomously optimize its own retrieval configuration based on failure logs. This closed-loop process allows the system to discover and refine effective retrieval strategies dynamically, significantly outperforming static baseline models on long-term memory benchmarks.

Orchard: An Open-Source Agentic Modeling Framework

AI Lab: Microsoft Research, Columbia University, UIUC

Summary: Orchard is an open-source framework built around a lightweight, Kubernetes-native environment service that decouples sandbox management from agent harnesses to enable scalable, cost-effective agentic training. The framework includes specialized training recipes for software engineering, web navigation, and personal assistant tasks, demonstrating state-of-the-art performance and strong generalization across diverse agent domains.

NEXUS: An Agentic Framework for Time Series Forecasting

AI Lab: Google, Pennsylvania State University

Summary: NEXUS is a multi-agent framework that improves time series forecasting by explicitly decomposing the task into macro-level trend projection, micro-level granular analysis, and dynamic synthesis of numerical and textual contexts. By separating these reasoning stages and incorporating a calibration loop, the system matches or exceeds the accuracy of dedicated numerical foundation models while providing highly interpretable forecasting rationales.

GLiGuard: Schema-Conditioned Classification for LLM Safeguard

AI Lab: Fastino Labs

Summary: GLiGuard is a highly efficient, 0.3B-parameter bidirectional encoder designed for real-time LLM content moderation by encoding task definitions and label semantics directly into the input schema. This non-autoregressive architecture evaluates prompt safety, response safety, harm categories, and jailbreak strategies simultaneously in a single forward pass, achieving competitive accuracy with models up to 90 times larger while drastically reducing inference latency.

🤖 AI Tech Releases

Interaction Models

Thinking Machines unveiled a research preview of Interaction Models: frontier models in which interactions is built in the model and not in the harness.

📡10 AI News You Need to Know About

1. Cerebras IPO debut : Cerebras priced its IPO at $185/share to raise $5.55B, and shares then soared 68% in their Nasdaq debut on May 14, pushing the AI chipmaker’s market cap to roughly $95B.

2. Recursive Superintelligence emerges from stealth: Former Salesforce chief scientist Richard Socher launched Recursive Superintelligence with $650M at a $4.65B valuation — led by GV and Greycroft with Nvidia and AMD Ventures participating — to build AI systems that recursively improve themselves through open-ended algorithms, starting with automating AI research itself.

3. OpenAI weighs suing Apple: OpenAI has hired an outside law firm and is considering legal action against Apple over the ChatGPT-Siri integration, which it claims has been buried in the UI and has fallen far short of projected subscription revenue.

4. Cisco cuts ~4,000 jobs while posting record revenue: Cisco is cutting roughly 5% of its workforce to redirect spending toward AI and cybersecurity, even as it touts record quarterly revenue and double-digit growth.

5. Adaption launches AutoScientist: Sara Hooker’s Adaption Labs unveiled AutoScientist, an automated fine-tuning system that co-optimizes data and model recipes and which the company says outperformed its in-house researchers by 35% (lifting win-rates from 48% to 64%), free for 30 days.

6. Musk v. Altman closing arguments: Lawyers for Musk and OpenAI delivered closing arguments in Oakland federal court on May 14, with the nine-person jury set to begin deliberating Monday on whether Altman and Brockman breached commitments to keep OpenAI a nonprofit.

7. xAI launches Grok Build: xAI rolled out an early beta of Grok Build, a new terminal-based coding agent and CLI for SuperGrok Heavy subscribers ($300/month), as Musk tries to close the coding-quality gap with Anthropic’s Claude.

8. Nebius Q1 sales up 684%: Nebius reported Q1 2026 revenue of $399M (up 684% YoY) and adjusted EBITDA of $129.5M, and announced it has secured up to 1.2 GW of power and land for a new owned AI factory in Pennsylvania.

9. Wispr in funding talks at $2B: Wispr AI, maker of the Wispr Flow voice dictation tool, is in talks for a ~$260M round led by Menlo Ventures that would value the company near $2B — nearly triple its $700M valuation from late 2025.

10. Former Alibaba Qwen lead launches new AI lab at ~$2B valuation : Junyang Lin, the former lead researcher behind Alibaba’s open-source Qwen models, is in talks with Chinese VCs Gaorong Ventures and HongShan to raise several hundred million dollars for a new AI lab at a roughly $2B post-money valuation — a level almost unheard of for a brand-new Chinese AI startup.

For today’s essay, I want to explore an idea that has become central to how we think about AI evaluations at LayerLens. This is not an essay about LayerLens, but about a simple and increasingly unavoidable thesis: evals are becoming the fourth pillar of modern AI, alongside compute, data, and models. As AI systems move from chatbots to agents, from demonstrations to production workflows, every meaningful task performed by every agent inside every company will need its own evaluation layer. Not generic benchmarks. Not leaderboard theater. Practical, dynamic, company-specific exams that measure whether an AI system can actually survive contact with real work. I call this idea Every Company’s Last eXam.

Humanity’s Last Exam is a very specific kind of artifact. It is what a field builds when the old report card stops working. The core observation behind it was simple: familiar benchmarks such as MMLU had become too easy for frontier systems to cleanly separate the leaders, so researchers assembled a harder, broader, multimodal test at the frontier of human knowledge, finalized at 2,500 questions after removing errors and questions that were too easily answerable with search. And then, almost immediately, the benchmark itself taught a second lesson: even “the last exam” needs maintenance. HLE-Verified later showed that noisy items and flawed answers could materially distort comparisons, and that systematic verification could shift measured accuracy by 7 to 10 percentage points on average. In other words, the benchmark was not a stone tablet. It was infrastructure.

That is the right analogy for where enterprise AI is going. Every company now needs its own last exam: a private, living evaluation suite that captures the highest-value, highest-risk, most context-heavy work its agents are supposed to perform. Not a generic IQ test for models. Not another public leaderboard. More like a company-specific CI system for cognition. The public benchmarks still matter, just as SPEC mattered for CPUs and ImageNet mattered for vision, but production truth has moved downstream into proprietary workflows, private documents, internal policies, odd exceptions, and all the sharp edges that never make it into a paper appendix. That is why top frontier labs now emphasize task-specific evals, production-derived datasets, continuous maintenance, and explicit definitions of success rather than vibe-based model selection.

The fourth pillar

Read more

There is a recurring fantasy in interpretability work, somewhere between a wish and an embarrassment. You stare at a residual stream activation — twelve thousand floats — and you want to ask it, in plain English, what are you thinking about? Sparse autoencoders give you a thousand sparse latents you then label by inspecting top-activating examples. Attribution graphs give you sprawling diagrams a researcher spends an afternoon parsing. Probes give you a yes/no. All useful. None of them talk back.

Anthropic’s new paper, Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations , is the first interpretability artifact in a while where the activation talks back. Literally. You point an NLA at a token in a Claude Opus 4.6 transcript and it produces a few bullet points of English describing what the model is thinking. That’s the deliverable. The paper is mostly an investigation of whether you should believe it.

The shape of the trick

Read more

💡 AI Concept of the Day: How State Space Models Went from Curiosity to Serious Transformer Competitor

There is this thing that happens in ML research where a line of work gets quietly good for years, and then one day you wake up and it’s suddenly competing with the dominant paradigm. State space models are having that moment right now.

For the past eight years, the transformer has been the only architecture that matters. Self-attention, key-value caches, next-token prediction — it’s all we think about. And for good reason: the thing works. But transformers have a problem that everyone in the field knows about and nobody has fully solved. Self-attention is O(n²) in sequence length. That’s not a theoretical concern anymore. When you’re trying to push context windows past a million tokens, or when you’re running inference on a 70B model and your KV-cache alone is eating 40GB of VRAM, quadratic scaling stops being an academic footnote and starts being the actual engineering bottleneck.

State space models offer a fundamentally different contract: linear time complexity, constant memory at inference, and no KV-cache at all. The question for the last three years has been whether they can match transformers on the things that matter — language modeling perplexity, in-context learning, reasoning. As of March 2026, the answer is: increasingly, yes.

Let me walk you through how we got here.

The Mathematical Foundation

Read more

Next Week in The Sequence:

• We continue our series about alternatives to transformers.

• In the AI of the week, we dive into Anthropic’s groundbreaking paper about natural language autoencoders.

• Our opinion section dives into an interesting idea: every company’s last exam.

Subscribe and don’t miss out:

📝 Editorial: Inside the Machine, Outside the Text Box

This week in AI had the strange texture of a market that is simultaneously becoming more scientific, more productized, and more speculative. The headlines looked disconnected at first: Anthropic published a fascinating interpretability paper, OpenAI released new voice models, SubQ made a controversial 12 million-token context claim, DeepSeek and Moonshot attracted enormous valuation attention, and Sierra raised at a level that would have sounded absurd for an AI customer service company only a few years ago. But underneath all of it is the same story: AI is moving from a model race into an infrastructure race.

Anthropic’s Natural Language Autoencoders paper was the most intellectually interesting development of the week. The idea is almost poetic: take the hidden activations inside a neural network and compress them into natural language, then try to reconstruct those activations from the explanation itself. In other words, language becomes a microscope for the model’s internal state. This is not a magical solution to interpretability. These explanations can be incomplete, noisy, or even misleading. But the conceptual shift matters. We are no longer just probing models with classifiers and activation maps; we are trying to build linguistic interfaces into the latent space. The model begins to explain itself in the medium humans understand best.

On the opposite side of the stack, OpenAI’s new voice model release pushed AI further toward becoming a native interface rather than a text box with better UX. Voice has always looked deceptively simple from the outside, but real-time speech agents require a brutal combination of perception, reasoning, latency management, interruption handling, emotional calibration, tool use, and memory. When this works, software changes shape. We stop “using an app” and start interacting with an operator. The difference is subtle but profound. Text-based AI feels like querying intelligence. Voice-based AI feels like being accompanied by it.

Then came SubQ’s controversial 12 million-token context announcement, the most provocative technical claim of the week. Long context has become one of the industry’s favorite flexes, but a native 12M-token window would represent something more than incremental progress. It would challenge the current architecture of retrieval-augmented generation, memory systems, chunking strategies, and agent orchestration. If models can directly absorb corpora at that scale, some of the scaffolding around AI applications starts to look temporary. Of course, claims like this demand skepticism. A massive context window is not the same as reliable reasoning over that context. But even the ambition is revealing: memory is becoming a frontier primitive.

The valuation news told the geopolitical and commercial version of the same story. DeepSeek and Moonshot are now being discussed at valuations that make them look less like startups and more like national AI infrastructure. Frontier model labs are increasingly priced as strategic assets: part software company, part cloud platform, part semiconductor leverage, part geopolitical option. The market is not merely valuing revenue; it is valuing position in the future computational order.

Sierra’s new valuation adds the enterprise counterpoint. While model labs chase frontier intelligence, Sierra is showing that applied agents can become enormous businesses by embedding directly into customer operations. The first trillion-dollar AI workflows may not look like science fiction. They may look like call centers, insurance claims, banking support, retail service, and enterprise processes slowly being rewritten around agents.

So the week’s lesson is clear: AI is becoming more inspectable, more conversational, more memory-rich, and more institutionally valuable. The race is no longer just about building smarter models. It is about building the interfaces, memory systems, deployment layers, and companies that turn intelligence into infrastructure.

🔎 AI Research

Natural Language Autoencoders: Turning Claude’s thoughts into text

AI Lab: Anthropic

Summary: This research introduces Natural Language Autoencoders (NLAs), a technique that translates complex language model activations into readable text to reveal a model’s internal, unverbalized reasoning. By applying NLAs during safety testing and model auditing, researchers can successfully detect when models secretly know they are being evaluated and uncover hidden misaligned motivations.

SkillOS: Learning Skill Curation for Self-Evolving Agents

AI Lab: UIUC, Google, and other institutions

Summary: This paper introduces SkillOS, an experience-driven reinforcement learning framework that enables self-evolving LLM agents to learn complex, long-term skill curation policies. By pairing a frozen agent executor with a trainable skill curator that updates and refines an external skill repository, SkillOS allows agents to effectively learn from sparse, delayed feedback, leading to more targeted skill usage and improved performance across diverse reasoning and multi-turn agentic tasks.

D-OPSD: On-Policy Self-Distillation for Continuously Tuning Step-Distilled Diffusion Models

AI Lab: The Hong Kong University of Science and Technology, Alibaba Group, University of California San Diego, The Chinese University of Hong Kong

Summary: This paper proposes D-OPSD, an on-policy learning paradigm for fine-tuning step-distilled diffusion models that leverages the inherited in-context capabilities of their LLM/VLM encoders. By assigning the model dual roles as both teacher and student with varying multimodal contexts, D-OPSD enables the learning of new concepts and styles without compromising the model’s original efficient few-step generation capabilities.

Agentic AI Systems Should Be Designed as Marginal Token Allocators

AI Lab: University of Illinois Urbana-Champaign

Summary: This position paper argues that agentic AI systems should be structured as economies that allocate marginal tokens based on a combination of quality, cost, latency, and risk rather than functioning merely as text generators priced by the unit. Adopting this marginal token allocation perspective helps explain and resolve recurring system failures—such as over-routing, over-delegation, and cache misuse—that arise when different layers of the AI stack are optimized in isolation.

COUNTING AS A MINIMAL PROBE OF LANGUAGE MODEL RELIABILITY

AI Lab: Stanford University

Summary: The authors introduce Stable Counting Capacity, a purely mechanical assay that tests a language model’s procedural reliability by having it count repeated symbols until failure, effectively removing semantic and knowledge-based confounds. Through extensive evaluation, the study reveals that current language models rely on finite, count-like internal states rather than open-ended logic, causing their procedural rule-following to collapse into guessing when these limited resources are exhausted.

Hallucinations Undermine Trust; Metacognition is a Way Forward

AI Lab: Google Research, Tel Aviv University

Summary: This paper reframes AI hallucinations as confident errors and argues that the inability of models to perfectly distinguish truths from errors creates an unavoidable tradeoff between utility and strict factuality. To overcome this stalemate, the authors propose developing metacognitive models capable of “faithful uncertainty,” which involves aligning a model’s linguistic uncertainty with its intrinsic uncertainty to preserve useful information while accurately communicating doubt to users.

🤖 AI Tech Releases

GPT-Realtime

OpenAI unveiled three new audio models to enable the construction of voice apps.

Gemma MTP

Google released Gemma Multi Token Prediction(MTP) , a new speculative decoding architecture that can predict multiple tokens at the same ti.

📡10 AI News You Need to Know About

1. DeepSeek targeting $45B valuation in first-ever funding round — DeepSeek is in talks for its first external venture round at a valuation that has reportedly jumped from $20B to $45B in weeks, led by China’s state-backed China Integrated Circuit Industry Investment Fund (the “Big Fund”) with Tencent and Alibaba reportedly in talks to participate, as founder Liang Wenfeng (who owns ~90% of the company) opens up the cap table primarily to issue employee equity and stem researcher poaching.

2. SpaceX ‘Terafab’ chip factory — SpaceX is considering spending an initial $55 billion (and up to $119 billion total) to build a multi-phase, vertically integrated semiconductor and advanced computing fab in Grimes County, Texas, with Tesla and Intel involved, to supply chips for AI servers, satellites, space data centers, and autonomous Tesla vehicles/robots.

3. Ethos $22.75M Series A — London-based Ethos raised a $22.75M Series A led by a16z (with General Catalyst, XTX Markets, Evantic, and Common Magic) to scale its voice-agent-powered expert network, which onboards roughly 35,000 experts per week and serves hedge funds, PE firms, AI labs, and consultancies.

4. QuTwo $380M valuation — Helsinki-based QuTwo, founded by ex-Silo AI CEO Peter Sarlin, raised a €25M (~$29M) angel round at a €325M (~$380M) valuation from a group of unicorn founders and Midas-listed investors to scale QuTwo OS, an orchestration layer for classical, hybrid, and quantum-inspired enterprise AI workloads.

5. SAP acquires Prior Labs / blocks rival agents — SAP announced plans to acquire Freiburg-based tabular foundation model startup Prior Labs (an “almost all-cash” deal) and invest €1B (~$1.16B) over four years to turn it into a European frontier AI lab for structured enterprise data, while simultaneously updating its API policy to block all third-party AI agents (e.g. OpenClaw) except SAP-endorsed ones like its own Joule and Nvidia’s NemoClaw.

6. CopilotKit $27M Series A — Seattle-based CopilotKit raised $27M (Series A + previously unannounced seed) led by Glilot Capital, NFX, and SignalFire to scale its open-source AG-UI protocol and launch CopilotKit Enterprise Intelligence, a self-hostable layer for embedding generative-UI AI agents inside enterprise apps used by customers like Cisco, Docusign, and Deutsche Telekom.

7. Sierra $950M raise — Bret Taylor’s Sierra raised $950M led by Tiger Global and GV at a post-money valuation north of $15B to expand its enterprise customer-experience AI agent platform, which the company says now serves more than 40% of the Fortune 50 and recently hit $150M in ARR.

8. Moonshot AI / Kimi $20B valuation — Beijing-based Moonshot AI is closing roughly $2B in new funding led by Meituan’s Long-Z (Dragon Ball) venture arm, with China Mobile and CITIC PE participating, at a post-money valuation above $20B, after Kimi’s annualized recurring revenue passed $200M in April.

9. Snap–Perplexity $400M deal terminated — Snap disclosed in its Q1 2026 investor letter that its $400M cash-and-equity partnership with Perplexity (announced last November to integrate Perplexity’s AI search into Snapchat’s Chat interface) “amicably ended” in Q1 after the two sides couldn’t agree on a path to broader rollout, with Snap’s 2026 sales guidance now assuming zero contribution from the deal.

10. Subquadratic / SubQ launch — Miami-based startup Subquadratic emerged from stealth on May 5 with $29M in seed funding (at a reported $500M valuation) led by Justin Mateen, Javier Villamizar, and others, claiming its first model SubQ 1M-Preview is the first LLM built on a fully subquadratic attention architecture (SSA) — with a 12M-token context window and a claimed ~1,000x reduction in attention compute versus frontier models.

For the first few decades of the internet, software had a reasonably stable assumption baked into it: the user was a human.

A human had eyes, a browser, a mouse, a password manager, a credit card, an email address, a tolerance for modal dialogs, and a finite amount of patience. The entire SaaS and consumer internet stack grew around this shape of user. Search engines ranked pages for humans. E-commerce sites optimized funnels for humans. CRMs tracked human sales reps selling to human buyers. Identity systems authenticated humans. Analytics systems measured human clicks, human sessions, human conversions.

Then we started building AI agents.

Read more

The interesting thing about NVIDIA’s new Nemotron 3 Nano Omni is not that it “does multimodality.” We already have a zoo of models that can caption images, transcribe speech, parse PDFs, answer questions about videos, and click around GUIs. The interesting thing is that Nemotron Omni is designed to make that zoo feel like a single animal.

Today’s multimodal agent stack often looks like a Rube Goldberg machine: audio goes to an ASR model, screenshots go to a VLM, PDFs are rendered into images or OCR’d into text, video gets sampled into frames, and then a language model tries to stitch the outputs together. Every boundary between models is a lossy compression step. The speech model may hear what was said but not what was on screen when it was said. The vision model may see the chart but not the voiceover. The planner gets a pile of summaries rather than a coherent sensory stream. Nemotron 3 Nano Omni is NVIDIA’s attempt to move the “eyes and ears” of an agent into a single efficient perception-and-reasoning model: video, audio, image, and text in; text out. NVIDIA announced it on April 28, 2026, positioning it as an open omni-modal reasoning model for agentic workflows like computer use, document intelligence, and long audio-video understanding.

Read more

💡 AI Concept of the Day: Return of the King: Unrolling the xLSTM Architecture

If you were training sequence models circa 2015, your entire mental model of the world was shaped by the Long Short-Term Memory (LSTM) network. Invented in the 1990s by Sepp Hochreiter and Jürgen Schmidhuber, the LSTM was the undisputed workhorse of deep learning. It translated our text, recognized our speech, and powered the first generation of Large Language Models.

Then came 2017. “Attention Is All You Need” dropped, and the entire AI ecosystem pivoted. We traded the deep, architectural elegance of the LSTM for the brute-force, highly parallelizable matrix multiplications of the Transformer. The Transformer won the hardware lottery because it allowed us to map the entire sequence onto a GPU grid and train it all at once.

Read more

Next Week in The Sequence:

• We continue our series about alternatives to the transformer architecture.

• In our opinion section, we are going to dive into the thesis of building software for AI agents instead of human consumers.

• The AI of the weeks dive into NVIDIA’s new Nemotron model.

Subscribe and don’t miss out:

📝 Editorial: Last Week in AI: The Great AI Fundraising Wars and a New Frontier Lab

This week in AI felt less like a product cycle and more like a sovereign debt auction for the future of cognition. The models are still improving, the demos are still fun, and yes, the agents are still occasionally lost in the hallway holding a JSON object. But the real story is that frontier AI is becoming an industrial-scale capital formation game.

The center of gravity was Anthropic. Reports this week said the Claude maker is weighing a new round that could value the company above $900 billion, potentially surpassing OpenAI’s most recent reported valuation and becoming the world’s most valuable AI startup. Reuters reported that Anthropic is considering offers at more than double its prior valuation, with TechCrunch adding that the round could be in the $40 billion to $50 billion range.

This is not just investor exuberance. It is the market trying to price a new kind of company: part model lab, part cloud tenant, part developer platform, part enterprise operating system. Anthropic’s recent strategic deals tell the story. Google committed up to $40 billion to Anthropic, while Amazon recently announced up to $25 billion in additional investment tied to a much larger cloud partnership. Anthropic has also been securing compute through Broadcom, CoreWeave, Amazon chips, and its own data-center ambitions.

The competition with OpenAI is now less about whose chatbot writes the better poem and more about who can assemble the stronger industrial stack. OpenAI, for its part, just renegotiated its Microsoft relationship to gain more flexibility to court Amazon and other cloud providers, while Microsoft keeps long-term economics through 2030. Reuters framed the change explicitly as helping OpenAI secure more compute and compete better with Anthropic in enterprise markets.

So the frontier race is mutating. The old question was: who has the best model? The new question is: who has the best capital structure, compute supply chain, cloud distribution, developer wedge, enterprise trust layer, and agent ecosystem? Intelligence is becoming infrastructure, and infrastructure has a balance sheet.

The same pattern is now showing up in vertical AI. Legal AI startup Legora extended its Series D by $50 million, bringing the total round to $600 million and reaching a $5.6 billion valuation, with Atlassian and Nvidia’s NVentures joining as investors. Legora has reportedly surpassed $100 million ARR, scaled to more than 1,000 organizations across 50 markets, and is positioning itself as an agentic operating system for legal work.

That makes the Harvey versus Legora battle one of the most interesting application-layer fights in AI. Harvey has the U.S. legal-tech aura, the elite law-firm mindshare, and an $11 billion valuation. Legora has the European-born insurgent energy, rapid ARR growth, and now corporate investors that understand workflow and compute. The punchline is not that lawyers will use chatbots. That is too small. The real punchline is that legal work is being decomposed into agentic workflows: research, drafting, diligence, review, negotiation, precedent search, risk analysis, and eventually entire deal rooms with AI copilots quietly moving the machinery.

Another signal came from London, where Ineffable Intelligence launched with a spectacular $1.1 billion seed round at a reported $5.1 billion valuation—Europe’s largest seed financing to date. Founded by former DeepMind researcher David Silver, one of the central figures behind AlphaGo, Ineffable is not pitching another chatbot, but a return to a more radical AI thesis: reinforcement-learning systems that learn through experience rather than merely compressing human-generated data. In a week dominated by Anthropic and OpenAI’s capital wars, Ineffable felt like the research-world version of the same phenomenon: frontier AI companies are now being funded not as startups, but as civilizational-scale experiments. The market is placing billion-dollar bets not only on who wins the current LLM race, but on what might come after it.

This week’s lesson is simple: AI is leaving the toy-box phase. Frontier labs are becoming trillion-dollar infrastructure companies. Vertical AI startups are becoming operating systems for professional work. The intelligence explosion may still be debated philosophically, but the capitalization explosion is already here.

🔎 AI Research

The Last Human-Written Paper: Agent-Native Research Artifacts

AI Lab: Orchestra Research, Stanford University, and others

Summary: This paper introduces the Agent-Native Research Artifact (ARA), a protocol that replaces traditional linear narrative papers with an executable, four-layer knowledge package optimized for AI agents. By preserving scientific logic, executable code, the branching exploration history, and grounded evidence, the ARA framework significantly improves an agent’s ability to extract knowledge, reproduce experiments, and extend prior research.

GLM-5V-Turbo: Toward a Native Foundation Model for Multimodal Agents

AI Lab: Z.ai & Tsinghua University

Summary: This technical report presents GLM-5V-Turbo, a multimodal foundation model designed to deeply integrate vision and language across perception, reasoning, planning, and execution. Through a novel CogViT vision encoder, Multimodal Multi-Token Prediction, and joint reinforcement learning over diverse tasks, the model achieves strong performance in multimodal coding and GUI agent tasks without sacrificing text-only capabilities.

Length Value Model: Scalable Value Pretraining for Token-Level Length Modeling

AI Lab: University of California, Santa Barbara, Apple Inc., and others

Summary: This paper proposes the Length Value Model (LenVM), which frames autoregressive generation length modeling as a token-level value estimation problem by predicting a discounted return for remaining generation steps. This annotation-free and scalable approach enables continuous control over the performance-efficiency trade-off during inference, allowing for highly accurate length constraint matching and generation horizon prediction.

Synthetic Computers at Scale for Long-Horizon Productivity Simulation

AI Lab: Microsoft

Summary: This paper introduces a scalable methodology to create realistic, user-specific synthetic computer environments populated with diverse directory structures and content-rich artifacts. By running long-horizon productivity simulations within these environments, the system generates rich experiential signals that significantly improve agent performance on both in-domain and out-of-domain productivity evaluations.

CO-DIRECTOR: Agentic Generative Video Storytelling

AI Lab: Google

Summary: This paper introduces CO-DIRECTOR, a hierarchical multi-agent framework that formalizes video storytelling as a global optimization problem to overcome semantic drift and cascading failures in current agentic pipelines. By combining multi-armed bandit optimization with a local multimodal self-refinement loop, the system dynamically explores creative trajectories and ensures consistent narrative and visual styles across sub-agents.

Recursive Multi-Agent Systems

AI Lab: UIUC, Stanford University, NVIDIA, MIT

Summary: This paper presents RecursiveMAS, a framework that scales multi-agent collaboration by casting the entire system as a unified latent-space recursive computation using a lightweight RecursiveLink module. Through an inner-outer loop learning algorithm, the system enables efficient cross-agent interaction and shared gradient-based credit assignment without text-based latency, leading to significant improvements in accuracy, inference speed, and token reduction across diverse benchmarks.

🤖 AI Tech Releases

Nemotron 3 Nano Omni Model

NVIDIA released Nemotron 3 Nano Omni Model, a multimodal, long-context model optimized for agentic tasks.

Claude Security

Anthropic announced Claude Security’s Beta, an addition to its enterprise solution focused on security vulnerability scanning.

Claude for Creative Work

Anthropic released Claude for Creative Work, a series of tools and connectors for creative and design platforms.

📡10 AI News You Need to Know About

1. Anthropic potential $900B+ valuation round — Anthropic is asking investors to submit allocations within 48 hours for an estimated ~$50B round at a targeted ~$900B valuation, expected to close within two weeks and likely to be its last private round before a planned IPO.

2. Legora hits $5.6B valuation — Swedish legal AI startup Legora extended its Series D by $50M, adding Nvidia’s NVentures and Atlassian as investors at a $5.6B post-money valuation after crossing $100M ARR, narrowing the gap with US rival Harvey. → Legora newsroom announcement

3. Flourish targets $2.5B valuation — Thomas Reardon, who led work on Meta’s Neural Band, is raising for a new energy-efficient AI startup called Flourish at a $2.5B valuation.

4. Parallel Web Systems hits $2B valuation — Parag Agrawal’s AI agent web-search/research API startup raised a $100M Series B led by Sequoia at a $2B valuation, just five months after its $740M Series A.

5. David Silver’s Ineffable Intelligence raises $1.1B — Former DeepMind reinforcement learning lead David Silver’s new UK-based AI lab raised $1.1B at a $5.1B valuation from Sequoia, Lightspeed, Nvidia, Google and others to build a “superlearner” that learns without human data.

6. China vetoes Meta’s $2B Manus deal — China’s National Development and Reform Commission ordered Meta and AI agent startup Manus to unwind their ~$2B acquisition, citing prohibition of foreign investment in the Manus project.

7. Amazon offers new OpenAI products on AWS — Days after OpenAI ended Microsoft’s exclusivity, AWS announced that Bedrock now hosts OpenAI’s latest models, Codex, and a new “Bedrock Managed Agents” service built on OpenAI’s reasoning models.

The next evolution of agentic SaaS isn’t more tool infrastructure. It’s a terminal.

I’ve been thinking a lot lately about why building agentic systems still feels so weirdly clunky, and I think I’ve finally put my finger on it.

Here’s the pattern that keeps repeating. You have an LLM. The LLM is, by any reasonable measure, the most capable text-processing machine ever built. It has read, in some statistical sense, every man page that exists, every Stack Overflow answer about awk, every dotfile checked into GitHub, every grumpy 2009 blog post about why your sed substitution didn’t work. It can produce a find ... -exec incantation that would make a 1990s sysadmin weep with nostalgia. It is, in a very deep sense, fluent in shell.

And then we sit it down in front of our shiny new SaaS platform and say: okay, we’re going to teach you how to use our product. Here are forty-seven carefully scoped MCP tools we wrote for you. Here is a JSON schema for each one. Here is a 200-page integration guide. Please do not improvise.

This is, I think, exactly backwards.

The thesis I want to argue is simple: the next phase of agentic SaaS is not about chat interfaces, and it’s not about ever-more-elaborate tool infrastructures. It’s about giving the agent a complete CLI and getting out of the way. Every SaaS, eventually, will ship a parallel command-line surface — not as a developer convenience, but as the primary interface for its non-human users. Which, increasingly, is most of its users.

The bitter lesson, applied to interfaces

Read more

DeepSeek’s releases always draw a lot of attention. Last week was the time for its v4 version.

The most interesting thing about DeepSeek-V4 is not that it supports a one-million-token context window. That number is impressive, but context length by itself is a poor proxy for intelligence. A model can accept a million tokens and still fail to use them. It can drown in KV cache, retrieve the wrong evidence, lose track of local syntax, hallucinate over compressed memory, or turn the entire prompt into a blurry statistical soup.

The real question is not: how much text can the model ingest?

The real question is: how much history can the model economically use?

DeepSeek-V4 is best understood as an answer to that question. It is not simply another frontier model release. It is a systems paper about making long-context reasoning practical. The model is designed around a simple but profound premise: million-token intelligence requires more than scaling the Transformer. It requires a new memory hierarchy, new attention mechanics, new training stabilizers, new optimizer choices, new quantization regimes, and a serving stack that can actually survive the economics of inference.

Read more

💡 AI Concept of the Day: The Unexpected Comeback of RNNs

If you were building sequence models circa 2015, your mental model of the world was entirely shaped by Recurrent Neural Networks (RNNs). There was a deep, architectural elegance to them. You feed the network a token, it updates a fixed-size hidden state, and it throws the token away. During inference, the memory footprint was beautifully constant—an $O(1)$ operation that could run efficiently on almost any hardware.

Then came 2017. “Attention Is All You Need” dropped, and the entire AI ecosystem pivoted. We traded the elegance of the RNN for the brute-force parallelizability of the Transformer. The Transformer won the hardware lottery because it allowed us to map the entire sequence onto a GPU grid and train it all at once. But we made a devil’s bargain: the Key-Value (KV) cache.

In a Transformer, the model must explicitly hold the high-dimensional representations of every previous token in memory to generate the next one. This is an O(N^2) operation. As we push models to 100K, 1M, and now multi-million token context windows, the compute graph becomes mathematically offensive. We are burning vast amounts of high-bandwidth memory simply doing memory reads.

This is why, if you watch the arXiv firehose closely right now, you will notice a massive vibe shift. We are witnessing the comeback of the RNN. But this is not a nostalgic return to the classic Long Short-Term Memory (LSTM) networks of the 2010s. The new generation of RNNs features larger states, data-dependent gating, and LLM-era training recipes. They are matching Transformer perplexity at scale, but keeping that sweet $O(1)$ inference cost.

Here is a look at the architectures driving the recurrent renaissance.

Read more

Next Week in The Sequence:

• Our series about transformer alternatives starts by exploring the comeback of RNNs.

• We dive into DeepSeek v4 and GPT 5.5.

• The opinion section we dive into an interesting thesis: a CLI for evertyhing

Subscribe and don’t miss out:

📝 Editorial: OpenAI Ships Agents, xAI Eyes Cursor, DeepSeek and Kimi Advance

This week in AI felt less like another cycle of model launches and more like a shift in the substrate of software itself. The important story is not simply that new models are becoming more capable. That has been the default trajectory for several years. The more interesting development is that models are becoming increasingly entangled with the systems where work actually happens: code editors, enterprise workflows, cloud environments, collaboration tools, and agentic interfaces.

OpenAI’s GPT-5.5 release is the obvious center of gravity. It represents the continued expansion of frontier-model capability across reasoning, coding, tool use, long-context work, and professional tasks. But the benchmark narrative is almost becoming secondary. A frontier model is no longer just a model. It is a runtime. It is the intelligence layer inside coding environments, research workflows, enterprise assistants, and autonomous systems. The model is becoming less like a smarter chatbot and more like a computational engine that can coordinate action.

OpenAI’s other releases made that thesis even clearer. Workspace Agents push ChatGPT from an individual productivity tool into a shared organizational substrate: Codex-powered agents that can live inside a company, run in the cloud, operate across tools like ChatGPT and Slack, follow permissions, remember context, and execute long-running workflows. This is not just “custom GPTs with enterprise packaging.” It is the beginning of AI as reusable institutional process. At the same time, ChatGPT Images 2.0 expands the surface area of AI work from language and code into visual production, with stronger text rendering, multilingual support, visual reasoning, and “images with thinking,” where the model can spend more time planning and refining before generating. Put together, these releases show OpenAI trying to make ChatGPT less like an app and more like a multimodal work environment: one place where text, code, images, tools, memory, approvals, and agents begin to converge.

The xAI deal with Cursor fits perfectly into this larger pattern. Cursor has become one of the clearest examples of AI-native software development moving from novelty to infrastructure. Code is the ideal environment for agents because it is explicit, testable, composable, and economically valuable. A coding agent can propose, edit, run, debug, and verify. It operates in a loop where progress can be measured. Whoever owns that loop owns one of the most important surfaces in the future of AI.

Meanwhile, DeepSeek V4 and Kimi 2.6 show how quickly the open and semi-open model ecosystem is compressing the frontier from below. The new competition is not merely about chat quality or leaderboard theater. It is about long context, coding performance, tool use, latency, cost, and agentic reliability. In other words, the battleground is shifting from intelligence as conversation to intelligence as execution.

This is the real theme of the week: AI is becoming operational. The model is no longer the product by itself. The product is the model plus the harness, the tools, the memory, the permissions, the environment, and the feedback loop. We are moving from models that answer questions to systems that perform work.

🔎 AI Research

Decoupled DiLoCo for Resilient Distributed Pre-training

AI Lab: Google DeepMind, Google Research

Summary: This paper introduces Decoupled DiLoCo, an evolution of the DiLoCo framework designed to improve the resilience of large language model pre-training against hardware failures and network issues. By separating compute across independent, asynchronously communicating "learners," the framework achieves significant improvements in training efficiency (goodput) while maintaining competitive model performance, even in highly fault-prone environments simulated through chaos engineering.

LLaDA2.0-Uni: Unifying Multimodal Understanding and Generation with Diffusion Large Language Model

AI Lab: Inclusion AI, Ant Group

Summary: This paper introduces LLaDA2.0-Uni, a unified discrete diffusion large language model that seamlessly integrates multimodal understanding and generation within a single framework. By discretizing visual inputs into semantic tokens and employing block-level masked diffusion, the model matches specialized vision-language models while supporting interleaved generation and reasoning.

SkillLearn Bench: Benchmarking Continual Learning Methods for Agent Skill Generation on Real-World Tasks

AI Lab: Carnegie Mellon University, Amazon AGI

Summary: The authors present SkillLearnBench, the first benchmark designed to evaluate continual learning methods for agent skill generation across 20 real-world tasks. Their evaluation reveals that while continual learning methods improve performance over no-skill baselines, they still fall short of human-authored skill levels and struggle with open-ended tasks.

Scaling Test-Time Compute for Agentic Coding

AI Lab: Meta Superintelligence Labs, University of Washington, New York University, Google DeepMind, Carnegie Mellon University, Princeton University

Summary: This paper proposes a test-time scaling framework for long-horizon coding agents by converting noisy rollout trajectories into compact, structured summaries. Utilizing Recursive Tournament Voting (RTV) for parallel scaling and Parallel-Distill-Refine (PDR) for sequential scaling, this representation-centric approach significantly boosts the performance of frontier models on challenging agentic benchmarks.

SWE-chat: Coding Agent Interactions From Real Users in the Wild

AI Lab: Stanford University

SWE-chat introduces the first large-scale dataset of real-world coding agent sessions, capturing over 6,000 interactions, 63,000 user prompts, and 355,000 tool calls from open-source developers. Analyzing this data reveals that while “vibe coding” is increasingly popular, it remains costly and introduces more security vulnerabilities, frequently prompting users to interrupt or correct the agent.

AutoAdapt: An Automated Domain Adaptation Framework for Large Language Models

AI Lab: Microsoft

Summary: AutoAdapt is an end-to-end automated framework designed to optimize the complex domain adaptation process for large language models under tight resource constraints. By employing a multi-agent debating system to navigate best practices and an LLM-based surrogate for efficient hyperparameter tuning, the framework achieves a 25% relative accuracy improvement over state-of-the-art automated baselines.

🤖 AI Tech Releases

DeepSeek v4

The new version of DeepSeek is here with 1M context length and impressive agentic capabilities.

Kimi 2.6

Kimi 2.6 launched with marquee capabilities in agentic coding.

ChatGPT Images 2.0

OpenAI released incredibly enhanced image generation capabilities in ChatGPT.

Workspace Agents

OpenAI unveiled Workspace Agents, a new experience for creating agents that can handle complex workflows inside ChatGPT.

ML Intern

Hugging Face open sourced ML Intern, an agent that researchs and write ML related code.

📡10 AI News You Need to Know About

• SpaceX preempted Cursor’s nearly-closed $2B funding round at a $50B valuation by offering a $60B post-IPO acquisition option — paying $10B as a “collaboration” fee in the interim — as the post-xAI-merger SpaceX scrambles to position itself as an AI company.

• Infosys and OpenAI announced a strategic collaboration combining Infosys Topaz Fabric with OpenAI’s Codex and frontier models to drive enterprise software engineering, legacy modernization, and DevOps automation at scale.

• NeoCognition, an AI agent lab founded by Ohio State professor Yu Su with co-founders Xiang Deng and Yu Gu, emerged from stealth with a $40M seed co-led by Cambium Capital and Walden Catalyst Ventures to build self-learning agents that develop world models of specific work environments.

• Anthropic took an additional $5B from Amazon (bringing total Amazon investment to $13B) in exchange for a $100B+ ten-year AWS commitment covering up to 5GW of Trainium2-through-Trainium4 capacity to train and serve Claude.

• Microsoft committed A$25B (~US$18B) to Australia by end-2029, expanding Azure AI supercomputing capacity by over 140%, deepening cyber defense work with the ASD, and pledging workforce-ready AI training for three million Australians by 2028.

• Jeff Bezos and Vik Bajaj’s physical-AI lab Project Prometheus closed a $10B round at a ~$38B valuation with JPMorgan and BlackRock among participants, while separately exploring up to $100B for a holding company to acquire industrial businesses whose operational data would feed back into the lab’s models.

• Bret Taylor and Clay Bavor’s Sierra acquired Paris-based, YC-backed Fragment — its third acquisition of 2026 after Opera Tech and Receptive AI — bringing co-founders Olivier Moindrot and Guillaume Genthial onto the team to anchor Sierra’s agent development efforts in France.

• ComfyUI raised $30M at a $500M valuation in a round led by Craft (with Pace Capital, Chemistry, and TruArrow), capitalizing on an open-source community of 4M users and 60,000+ nodes that has made its node-based workflow the de facto control layer for production-grade generative media.

• Google committed up to $40B in Anthropic — $10B now at a $350B valuation, with $30B more contingent on performance milestones — alongside an expanded Google Cloud arrangement providing 5GW of TPU-based compute over the next five years.

• Meta signed a multi-year, multi-billion-dollar deal to bring tens of millions of AWS Graviton5 cores into its compute portfolio for agentic-AI inference workloads, becoming one of AWS’s largest Graviton customers and validating the thesis that agentic AI is shifting demand back toward CPUs.

The most important question in agentic software is not “Which model?” It is “What can the model touch?”

A language model by itself is a strange kind of brain in a jar. It can predict, plan, summarize, and improvise, but it has no hands. The moment we give it tools, it becomes something else: not merely a chatbot, but an operator. It can read files, write code, open issues, call APIs, move tickets, delete emails, deploy infrastructure, or wake you up at 3 a.m. because a background workflow misread a calendar event.

So the real primitive of agentic systems is the interface between the model and the world.

Two candidates have emerged as the main bridge: the command-line interface, or CLI, and the Model Context Protocol, or MCP. They represent two different philosophies. CLI says: “The best tool interface already exists. It is the Unix process. Text in, text out, exit code, compose everything.” MCP says: “Agents need structured, discoverable, typed tools. Give them a protocol, schemas, resources, prompts, permissions, and a client-server architecture.”

Read more

Claude Opus 4.7 shipped last week. The benchmarks are what you’d expect from a two-month incremental release — SWE-bench Verified 87.6%, SWE-bench Pro 64.3%, MCP-Atlas +14.6pp, state-of-the-art on GDPval-AA for economically valuable knowledge work, XBOW visual-acuity 54.5% → 98.5%, finance and document reasoning up, BrowseComp and long-context multi-needle retrieval down. Fine. Worth skimming. But the raw numbers undersell what actually changed.

The easier way in is to look at what got removed from the API, because the release is as much about the contract between you and the model as it is about the weights.

If you migrate a 4.6 harness to 4.7 and it still sets temperature, top_p, top_k, or thinking.budget_tokens, you get a 400. Not deprecated with a warning — gone. The only supported thinking mode is adaptive. In their place: an effort enum (low, medium, high, xhigh, max) and task_budget, a soft token ceiling the model can actually see. Every one of the removed parameters was a sampling-level control — you were reaching into the decoding loop and fiddling with token probabilities. What replaces them are semantic controls. You’re no longer tuning the softmax; you’re telling the model how hard to think and how much runway it has.

That’s basically the release in one sentence. The inference-time interface has shifted from stochastic sampling knobs to self-paced budgets, and the model has been trained to sit inside that interface responsibly. Everything else — self-verification, the literal instruction following, 1:1 pixel mapping, file-system memory, differential capability shaping — is downstream of this posture. Let me walk through what the new contract actually buys you.

Self-verification as a trained behavior, not a prompt trick

Read more

💡 AI Concept of the Day: Beyond Transformer: A New Series

If you have been watching the arXiv firehose lately, you can feel a very palpable vibe shift. Today, we are starting a new series to map out exactly what is happening: the search for novel alternatives to the Transformer architecture.

For the better part of a decade, the entire artificial intelligence ecosystem has essentially been a giant, spectacularly funded wrapper around a single mathematical operation: self-attention. The Transformer won the hardware lottery of the late 2010s. It was beautifully parallelizable across GPUs, and its mental model was intuitively simple—every token looks back at every previous token to decide what to do next.

Read more

Next Week in The Sequence:

• We start a new series about alternatives to transformers including text diffusion models, SSMs and many more.

• The opinion installment will discuss a hot topic: MCP or CLIs?

• The AI of the week section dives into Claude Opus 4.7.

Subscribe and don’t miss out:

📝 Editorial: Last Week in AI: Anthropic and OpenAI Enter a New Phase

This week brought a particularly interesting cluster of releases from Anthropic and OpenAI. Anthropic pushed on both the model and product fronts with Claude Opus 4.7 and the new Claude Design, while OpenAI expanded in two different directions with GPT-Rosalind and the latest Codex. Put together, these launches say something important about where the frontier is heading. The story is no longer just about smarter chatbots. It is about AI splitting into distinct product forms: the general-purpose reasoning model, the domain specialist, and the workflow-native agent.

Anthropic’s side of the week was especially revealing because it showed both ends of that transition. Claude Opus 4.7 looks like a refinement of the frontier generalist: stronger for advanced software work, better at following instructions, and more reliable in longer chains of reasoning. But Claude Design is the more interesting signal. It packages Claude not just as a model that can describe visual work, but as a collaborator for actually producing it—designs, prototypes, slides, and one-pagers. That matters because it turns design from a prompting exercise into a workflow. Anthropic is not just improving a model; it is carving out a new interface category where reasoning and visual production start to merge.

OpenAI’s releases tell a parallel story, but from a different angle. GPT-Rosalind represents the rise of the specialist model: a system built specifically for biology, drug discovery, and translational medicine rather than a general model stretched into a scientific costume. That is a meaningful shift. The next wave of AI advantage will not come only from one model doing everything reasonably well. It will come from systems that are deeply adapted to the language, tools, and constraints of high-value domains. Rosalind suggests that science is becoming one of the first places where that specialization is explicit, productized, and strategically central.

Then there is Codex, which may be the clearest signal of all. The newest version is not just about code generation. It can operate a computer, use tools and apps, remember preferences, connect to remote environments, and handle ongoing work over time. That changes the meaning of “coding assistant.” Codex is inching toward something broader: an agentic operating layer for software and knowledge work. The important shift is not that it writes better code. It is that it increasingly participates in the surrounding workflow—browsing, coordinating, iterating, and executing across environments rather than waiting passively for the next prompt.

The deeper pattern across Anthropic and OpenAI is that frontier AI is fragmenting into real products. Anthropic is pushing the frontier generalist into adjacent creative workflows. OpenAI is pushing both specialization and agentic execution. And together they are making the competitive landscape much more interesting. The real race is no longer just who has the smartest model on a benchmark. It is who can turn intelligence into the most compelling system for actual work. That is what made this week’s releases worth watching.

🔎 AI Research

Introspective Diffusion Language Models

AI Lab: Together AI, University of Illinois Urbana-Champaign, The University of Texas at Austin, Princeton University, and Stanford University

Summary: Diffusion language models traditionally lag behind autoregressive models because they lack introspective consistency, meaning they do not reliably agree with their own generated tokens. To address this, the authors introduce the Introspective Diffusion Language Model (I-DLM), which utilizes a novel strided decoding algorithm to simultaneously verify and generate tokens, matching autoregressive model quality while significantly improving serving efficiency.

Efficient RL Training for LLMs with Experience Replay

AI Lab: FAIR at Meta and NYU Courant Institute and CDS

Summary: This paper challenges the prevailing assumption that large language model reinforcement learning requires strictly fresh, on-policy data by demonstrating the effectiveness of experience replay. Through a systematic study, the authors show that a well-designed replay buffer can drastically reduce expensive inference compute costs during training while maintaining or even improving the model’s final performance and output diversity.

Lyra 2.0: Explorable Generative 3D Worlds

AI Lab: NVIDIA

Summary: Generating large-scale 3D scenes using video diffusion models often suffers from spatial forgetting and temporal drifting over long camera trajectories. Lyra 2.0 mitigates these degradations by utilizing per-frame 3D geometry for history retrieval and self-augmented training to correct drift, enabling the creation of persistent, explorable 3D worlds from a single input image.

Parcae: Scaling Laws For Stable Looped Language Models

AI Lab: University of California, San Diego and Together AI,

Summary: Looped architectures offer a way to increase a model’s compute without increasing its parameter footprint by repeatedly passing activations through a block of layers, but existing training methods for these models are often unstable. To resolve this, the authors propose Parcae, a novel, stable looped architecture that constrains the spectral norm of its injection parameters, enabling the discovery of predictable scaling laws for both training and test-time compute that improve model quality while keeping parameter counts fixed.

ELT: Elastic Looped Transformers for Visual Generation

AI Lab: Google DeepMind

Summary: Conventional visual generative models rely on deep stacks of unique transformer layers, whereas this paper proposes Elastic Looped Transformers (ELT) that use iterative, weight-shared blocks to drastically reduce parameter counts. By utilizing an Intra-Loop Self Distillation (ILSD) training method, ELT achieves high-fidelity image and video generation while enabling Any-Time inference, allowing users to dynamically trade off computational cost and generation quality at test-time.

Many-Tier Instruction Hierarchy in LLM Agents

AI Lab: Johns Hopkins University

Summary: Current instruction hierarchy paradigms use a small, fixed set of privilege levels, which is inadequate for real-world agents receiving conflicting instructions from numerous heterogeneous sources. To address this limitation, the authors introduce the Many-Tier Instruction Hierarchy (ManyIH) and an accompanying benchmark, MANYIH-BENCH, revealing that current frontier models perform poorly when resolving conflicts across arbitrarily many privilege levels.

🤖 AI Tech Releases

GPT-Rosalind

OpenAI released GPT-Rosalind, a new frontier reasoning model for life science research.

Gemini on Mac

Google announced a release of the Gemini app for macOS.

Codex for (Almost) Everything

OpenAI unveiled Codex for (almost) everything, a new agents that extends

Chrome Skills

Google released Chrome Skills to integrate AI workflows.

Claude Design

Anthropic Labs unveiled Claude Design, a collaborative experience to create visual artifacts.

Qwen3.6-35B-A3B

Alibaba Qwen open sourced Qwen3.6-35B-A3B, a more efficient version of its marquee agentic coding model.

📡AI News You Need to Know About

• Cursor in talks at $50B valuation — Four-year-old Cursor is nearing a $2B+ round led by returning backers Thrive and a16z at a $50B pre-money valuation (with Battery and Nvidia also expected to participate), as it projects ending 2026 at a $6B+ ARR run rate.

• OpenAI’s former product chief and Sora head leave — Kevin Weil (ex-CPO, most recently running the OpenAI for Science team, which is being decentralized into other research groups) and Bill Peebles (the researcher behind Sora) both announced their departures on Friday as OpenAI sheds “side quests” to focus on enterprise AI and its forthcoming superapp — with enterprise CTO Srinivas Narayanan reportedly leaving as well.

• Cerebras files publicly for US IPO (second attempt) — AI chipmaker Cerebras Systems publicly filed its S-1 with the SEC on Friday, disclosing $510M in 2025 revenue and $87.9M of net income (vs. a $484.8M loss on $290M of revenue in 2024), and plans to list on Nasdaq as “CBRS” with Morgan Stanley, Citi, Barclays, and UBS leading the offering.

• Factory hits $1.5B valuation — Enterprise AI coding startup Factory raised a $150M Series C led by Khosla Ventures at a $1.5B valuation, adding Keith Rabois to its board and bringing customers like Morgan Stanley, EY, and Palo Alto Networks onto its “Droids” agent platform.

• Upscale AI at $2B valuation — Seven-month-old AI infrastructure startup Upscale AI is reportedly in talks to raise $180M–$200M at a ~$2B valuation, its third round since launching, with Tiger Global, Xora, and Premji Invest as existing backers despite no shipped product yet.

• Hightouch hits $100M ARR — Hightouch told TechCrunch it has reached $100M ARR — $70M of which came in the 20 months since launching its brand-aware AI marketing product used by Domino’s, Chime, PetSmart, and Spotify. → TechCrunch exclusive; no earlier original source found.

• Gitar exits stealth with $9M — Gitar, founded by ex-Uber/Google/Intel engineers Ali-Reza Adl-Tabatabai and Gautam Korlam, emerged from stealth with a $9M seed led by Venrock (with Sierra Ventures) to build AI “agentic quality gates” for code review and CI validation.

• TSMC Q1 profit beats estimates — TSMC reported Q1 2026 net income of NT$572.5B (up 58% YoY) on revenue of US$35.9B (up 40.6% YoY in USD terms), guided full-year revenue growth above 30%, and said Middle East conflict has not dented AI chip demand.

• Accel raises $5B for late-stage AI bets — Accel announced $5B in fresh capital — $4B for its fifth Leaders Fund (targeting 20–25 ~$200M late-stage checks in AI, robotics, defense, and data center infrastructure) plus a $650M LP sidecar — after Anthropic and Cursor marks soared.

• Jane Street invests $1B in CoreWeave, commits $6B in cloud spend — Jane Street committed ~$6B to use CoreWeave’s AI cloud (including NVIDIA Vera Rubin capacity) and separately bought $1B of Class A stock at $109/share, bringing total commitments to $7B.

There is a meaningful difference between getting a model to write code and getting a model to reliably build software. The first is a neat demo. The second is a new engineering discipline. We are talking about harness engineering.

The idea is simple, but the implications are deep. Instead of treating the model as a magical coding oracle, you treat it as a powerful but imperfect operator inside a carefully designed environment. The goal is no longer to write the perfect prompt. The goal is to build the surrounding system so that good behavior becomes easy, bad behavior becomes visible, and failure becomes recoverable.

This is an important shift because most teams still think of agentic software as an interface problem. They focus on instructions, tone, and phrasing. But once an agent is doing meaningful work over long horizons, the bottleneck is rarely language alone. It is structure. It is visibility. It is memory. It is validation. It is the quality of the rails surrounding the model.

In other words, the real product is not the prompt. It is the harness.

OpenAI recently gave a useful name to a pattern many of us have been discovering the hard way: harness engineering. In its post on the subject, the company makes a strong argument that the real challenge is no longer just getting models to generate code, but building the surrounding environment—tools, constraints, plans, observability, documentation, and feedback loops—so agents can operate reliably inside production systems. That framing is exactly right, and it is a useful place to begin.

But this essay is not a restatement of OpenAI’s idea. I want to use that framing as a starting point and blend it with my own lessons from working with agentic systems in practice. The interesting part of harness engineering is not the label itself. It is the collection of non-obvious truths that appear once you move beyond one-shot demos and start asking agents to do real work over long horizons. At that point, the bottlenecks stop looking like prompt problems and start looking like engineering problems: memory, visibility, verification, architecture, process, and recovery.

And that is where the real lessons begin.

Lesson 1: When Agents Fail, the Environment Is Usually Underbuilt

Read more