More on our distillation series.
To keep you current, the frontier update section will provide mini deep dives about the new DeepSeek and GLM model as well as NVIDIA’s Lighting and Switchyard releases.
Will discuss some robotics stacks you need to track.
The opinion section, will discuss some ideas to help you understand the financing structures that are taking place in AI compute.
There was a time when following AI was relatively simple. A new model appeared, someone posted a benchmark table, and we updated the leaderboard in our heads.
That mental model is rapidly becoming obsolete.
Consider what happened this week. SpaceX officially closed its $60 billion acquisition of Cursor, one of the defining products of the AI coding era. At almost the same time, SpaceXAI released Grok 4.6, a model explicitly optimized for long-running agents, coding, and multi-step knowledge work. The important detail is not that Grok moved a few points on a benchmark. It is that Grok now flows directly into Cursor, Grok Build, GitHub Copilot, APIs, and autonomous agents.
The model is becoming the stack.
Think of the early cloud era. AWS did not win because EC2 had the prettiest virtual machine. It won because compute became attached to storage, databases, networking, identity and eventually an enormous developer ecosystem. Intelligence appears to be following the same path.
Anthropic seems to understand this. The company is reportedly discussing a roughly $6 billion acquisition of Decart AI, which works on model infrastructure, world models and compute optimization. The deal is not finalized, but the direction is interesting: one of the strongest model companies is reaching down the stack toward the machinery required to produce intelligence more efficiently.
Meanwhile, the frontier itself keeps getting more crowded.
China’s Z.ai announced GLM-5.3, showing surprisingly strong cybersecurity capabilities and again demonstrating how quickly open-weight models are compressing the gap with closed systems. If the first phase of the AI race was about discovering how to build frontier models, the second may be about how quickly everyone else can reproduce the recipe.
And then there is River AI, founded by former xAI co-founder Igor Babuschkin, which raised an extraordinary $1.1 billion this week. River’s thesis is almost the mirror image of the giant labs: instead of renting intelligence from one enormous generic model, companies and individuals should train models on their own data, rewards and preferences—and ultimately own the resulting intelligence. Its API already exposes fine-tuning and reinforcement learning across open models.
This creates an interesting tension.
One future looks vertically integrated: compute → model → agent → application → user.
The other looks modular: open model → proprietary data → reinforcement learning → personalized intelligence.
Both are racing toward the same scarce resource: not GPUs, parameters or even tokens, but feedback loops.
Now onto the most important AI developments of the week.
AI Lab: Microsoft
Summary: This paper introduces a full-bandwidth transformer that utilizes latent feedback decoding to fuse the previous top-layer hidden state with the current token embedding, thereby widening the model’s vertical communication channel. Trained via a scheduled multi-pass objective, this architecture matches or exceeds the performance of standard transformers trained on up to 1.5× more data, improving reasoning and coding generation with negligible inference overhead.
AI Lab: Salesforce AI Research
Summary: This research frames LLM agent self-improvement as a natural selection process across a population of agent harnesses (prompts, tools, and control flows), allowing for continuous capability evolution while keeping the base model weights completely frozen. By relying on a strict preserve-and-extend contract and measured fitness from task verifiers, the system effectively discovers and merges complementary skills without regressing on previously solved tasks.
AI Lab: University of Illinois Urbana-Champaign, Google
Summary: This paper introduces the Promptable Gaze Target Estimation (PGE) task and the GazeAnywhere model, which shifts gaze analysis to an end-to-end, concept-driven framework conditioned on text or visual prompts rather than relying on brittle, multi-stage pipelines. By simultaneously handling subject localization, in-frame presence, and gaze target heatmap estimation, GazeAnywhere achieves state-of-the-art results on multiple benchmarks, including a challenging real-world clinical dataset.
Summary: This paper unconditionally proves that at least two-thirds of the nontrivial zeros of the Riemann zeta function are simple and lie on the critical line, significantly improving upon previous unconditional records. The author achieves this by replacing the Riemann hypothesis’s conditional positivity requirement with a rank-trace inequality applied to a finite compression of Weil’s Hermitian form, and the findings are formally verified using Lean 4.
AI Lab: Anthropic
Summary: This research explores the coordination and behavior of multiple AI agents working together, demonstrating that while swarms can effectively tackle complex tasks like software vulnerability detection, they also exhibit distinct failure modes such as high conformity and rapid collusion. The study emphasizes the urgent need to understand these systemic risks as autonomous agent-to-agent interactions scale to potentially exceed human interactions in real-world environments.
AI Lab: Google Research, Technion – Israel Institute of Technology
Summary: This paper introduces a behavioral framework and the WikiProfile benchmark to evaluate whether factual errors in large language models stem from missing knowledge (”empty shelves”) or inaccessible encoded facts (”lost keys”). By analyzing over 4 million responses, the authors demonstrate that while encoding is nearly saturated in frontier models, recall remains the primary bottleneck, though inference-time computation (”thinking”) can effectively recover a substantial portion of these otherwise inaccessible facts.
Grok 4.6 xAI’s new frontier model for coding and agentic work landed on the API with a 500K context window, $2/$6 per million tokens below 200K prompt tokens, and a new xhigh reasoning effort level on top of low/medium/high.
Z.ai shipped GLM-5.3 with the tagline “Built to Code. Ready for Cyber Defense,” built entirely through post-training on the same 743B base as GLM-5.2, and available now via GLM Coding Plan and ZCode with API access and open weights staged behind safety evaluations.
NVIDIA released a 30B MoE with 3B active parameters on a hybrid Mamba-2 + MoE + attention architecture with a 1M context window, under the permissive OpenMDW-1.1 license, paired with an open-source routing library that sends each step of an agent workflow to the cheapest capable model.
The V4 Pro flagship left preview and went GA across app, web, and API with no calling-method change, positioned squarely on agent capability, alongside native OpenAI Responses API support and three thinking-effort levels for both Pro and Flash.
Databricks closed a $5 billion round at a $190 billion valuation led by Coatue, after crossing a $7 billion revenue run rate with more than 80% year over year growth in Q2, though Ghodsi told TechCrunch he only wanted $1 billion and saw $15 billion of investor interest.
SpaceX completed its $60 billion all-stock acquisition of Anysphere, issuing about 389 million Class A shares and folding Cursor into its SpaceXAI division as a wholly owned subsidiary.
Anthropic is reportedly in talks to acquire Decart for about $6 billion, a deal that would bring the Israeli startup’s chip-efficiency stack and world models into Anthropic’s inference and performance org ahead of a rumored IPO.
Lovable raised a $400 million Series C at a $13.3 billion valuation led by Menlo Ventures and the EQT-managed Scaleup Europe Fund, doubling its December mark as ARR tracks toward $600 million.
Cognition is in early talks to raise more than $1 billion at a $40 billion valuation, less than three months after its $26 billion round, with annualized revenue approaching $1 billion.
OpenAI acquired NextSlide, a roughly year-old startup that turned prompts and documents into editable decks, with founder Ahmed Beshry and team now working on ChatGPT and terms undisclosed.
River AI, the two-month-old startup from xAI co-founder Igor Babuschkin, raised $1.1 billion across seed and Series A led by General Catalyst and AMP PBC, with strategic money from NVIDIA and AMD Ventures, to build a full stack for personally owned models.
IBM announced a strategic partnership with OpenAI that embeds GPT-5.6, Codex, and ChatGPT Work into IBM Consulting Advantage and stands up a dedicated OpenAI practice with thousands of certified consultants.
A Thrive Capital letter to LPs revealed its $516 million 2022 early-stage fund is now marked above $3.7 billion on OpenAI and SpaceX positions, and that the firm is selling part of its OpenAI stake.
CoreWeave reported Q2 revenue of $2.58 billion, up 112% year over year, with revenue backlog around $104 billion and full-year guidance raised to $12.4 billion to $13.2 billion.