Yesterday covered an Anthropic IPO size target and Codex opening iMessage. This morning the story is a harness, not a ticker: Nvidia’s August 21 AVO write-up, plus a Poolside investor letter first reported by Newcomer.

The model on the scoreboard is Claude. The system that finished the games is not.

Lead: AVO did not ship a new brain. It shipped a loop that remembers.

Nvidia’s technical blog (Aug 21) is the primary. Agentic Variation Operators — AVO — completed all 183 levels across the 25-environment ARC-AGI-3 public set at 100.00 RHAE, in 6,624 environment actions. Same Claude Opus 5 backbone. ARC Prize separately scores that model at about 30% at High reasoning effort.

The authors put the thesis in the first paragraph: a frontier language model is only one component of an agent. The harness decides how context arrives, how tools fire, what state survives, and whether a stalled loop gets redirected.

AVO was not built for puzzle games. It was built to evolve CUDA kernels. In the attention-kernel study it ran seven days, explored more than 500 directions, committed 40 kernel versions, and beat cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% on DGX B200 boxes. Then the team swapped the task interface, kept the agent loop, and pointed the same machinery at ARC-AGI-3 as text-only 64×64 grids. No images. No rule sheet. No stated goal.

Bold insight: treat the 30-to-100 gap as a system result, not a model upgrade. Nvidia says this is not a controlled ablation. VISTA and AVO differ on backend, observation (VISTA’s primary config is a 512×512 PNG; AVO sent a text grid), memory, and context management. The 100.00 is public set only, on Nvidia’s own reimplementation of the task interface, not the semi-private or private competition sets, and not an official ARC Prize leaderboard entry. The paper is on arXiv as 2603.24517.

That is still the operator story of the week. If your SWE-agent dies after hour two, you do not need a newer checkpoint first. You need persistent memory that carries failed attempts forward and a supervisor that notices you are re-trying the same dead kernel. OfficeChai has the cleanest secondary write-up of those caveats.

If you are actually running long loops this weekend, the boring kit still wins: a USB-C dock so the box does not brown out mid-eval, a second monitor for the supervisor log, and an external SSD for the memory store you keep pretending lives in the chat.

Secondary: six billion for the factory, one billion for the stub

Same 24-hour window, different layer of the stack. PYMNTS, citing Newcomer and The Information: Nvidia will pay 6 billion for a non-exclusive license to Poolside’s Model Factory — the platform behind the open-weight Laguna coding models — and inject 1 billion at a 12 billion pre-money valuation. Job offers go to 109 Laguna staff. The three co-founders stay. The investor letter’s line is the one to keep: not an acquisition and not an acquihire.

Bloomberg has the same term sheet. The Next Web notes the license fee is expected to flow to existing backers by the end of 2027, and that this is the third time Nvidia has used the Groq / Enfabrica shape: license the IP, hire the bench, leave a company on the letterhead.

Poolside told investors that staying independent would have required more Nvidia hardware than it could get. Laguna is a real coding-model line, not vapor — Laguna M.1 has been cited around 72.5% SWE-bench Verified — but it is not the frontier. Nvidia is not buying a leaderboard. It is buying a model factory and the people who already know how to run it on Nvidia silicon.

Bold insight: yesterday’s IPO rumor was raise-size theater. Today’s deal is how the chip vendor absorbs a coding-model lab without a merger filing. Same subtractive move as the Groq license: keep the stub independent, take the factory.

If you are still deciding whether to fine-tune locally or rent the API, a workstation GPU and a NVMe enclosure are the cheaper way to learn what a “model factory” actually costs before someone invoices you in the billions.

Tertiary: DeepSeek grew eyes. Gemma hit a billion.

DeepSeek’s Aug 21 changelog shipped V4-Flash-Vision-Exp on the paid API. Text agents, reasoning, and world knowledge stay on par with V4-Flash. Multimodal agent benches jump; SiliconANGLE reports it beating Anthropic Opus 4.8 on ALE (1,000+ multi-step app tasks) and ZeroBench (100 hard image-analysis tasks). Images bill at up to 384 tokens each at Flash pricing. Files API is free: upload once, reuse by file_id. Model string: deepseek-v4-flash-vision-exp. DeepSeek Harness 0.1.1 dropped the same day.

This is still API-only and experimental. Do not wait for weights this weekend. The useful comparison is architectural, not tribal: Nvidia just showed a text-grid agent beating a vision harness on ARC, and DeepSeek just shipped vision so its cheap Flash stack can see the desktop. Different bets on where the pixels should live.

Same open-weight tape, already aging: Google says Gemma passed 1 billion downloads and 100,000 community variants. A DevRel note: that count excludes Android and Chrome. Qwen claimed 3 billion five days earlier. Download counters are not usage. If you want a local box that actually runs a Gemma variant this week, a Raspberry Pi 5 is still the cheapest always-on node.

Takeaway

1. Today: read the Nvidia post as a harness paper, not an ARC solve. Public set. Own interface. No private-set claim.
2. This week: if your agent loops are stalling, add a supervisor and a memory file before you swap models. The 30-to-100 gap is the argument.
3. Capital: Poolside is a non-exclusive factory license plus a 109-person bench transfer. Watch the Groq-shaped filings, not the Laguna leaderboard.

The spaghetti this morning is a coding agent that finished 183 games it was not trained for, a chip vendor that licensed the factory instead of buying the company, and a cheap Flash model that just learned to see. Scoreboards still matter. Which loop you wrap around the model matters more.

Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.

Some links may be affiliate links. If you’re buying hardware to run local models, this affiliate link helps keep the lights on.