Deep Learning Weekly: Issue 477
Deep Learning Weekly Miko Planas
Claude Haiku 5.5, new Meta Muse models, and Anthropic’s source control rebuild all landed this week. The real story: models are getting cheaper, faster, and weirder about their own context.
Based on reporting by Deep Learning Weekly, Miko Planas — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
This week’s Deep Learning Weekly reads like a snapshot of where the field is headed: smaller models trying to look bigger, bigger systems trying to get less fragile, and a steady push to make agents do more of the messy work themselves. Claude Haiku 5.5 leads the first camp. It’s priced at 0.10/0.50 per million tokens below 100,000 context and posts 39.2% on Terminal-Bench 4.0 at max effort, well ahead of GPT-6 Luna’s 16.4% in the source’s framing.
Meta’s new Muse Wonder models push on performance, but not equally on access. Muse Wonder Full hits 56 on the Artificial Analysis Intelligence Index, yet it ships subscription-only. The API-available Muse Wonder reaches 52 and costs 4/20 per million tokens. That split says a lot about how frontier model releases now work: the most capable version is not always the one developers can actually plug into their products.
Cloudflare’s Clef line is moving the other direction, widening from decision models into native audio, video, and image input on a Qwen3-Omni-30B-A3B base. The company says the system returns text decisions at a median of about 130ms. Elsewhere, TypeSafe closed an $870 million round at a $7.5 billion valuation, led by a16z with Sequoia and DCVC, under a month after its Jev decision model launched. That is a very large vote of confidence for a company still fresh off a model debut.
The software plumbing keeps getting more serious too. One guide walks through splitting a single LLM across two GPU nodes on EKS with separate prefill and decode workers behind one OpenAI-compatible endpoint. Another shows how browser agents can pay for x402-protected resources while AgentCore handles budgets and signs transactions without exposing wallet credentials. And Anthropic’s source-control rewrite stands out for its candor: the company replaced git with an in-house system in five months using seven people after merges climbed from 2,500 to more than 30,000 per day.
On the research side, the common thread is control. SELFCOMPACT lets a model decide when to compact its own long traces instead of forcing fixed-interval summarization, and the paper claims it matches or beats the baseline while using 30–70% less per-question cost. MASS takes a more ambitious swing at recursive self-improvement by having a single base model propose, execute, and judge multi-agent workflows; over two cycles with Qwen3.6-27B, it posts 1.2-1.6x higher performance per output token on four open-ended public benchmarks. The field keeps asking models to be their own janitors, editors, and managers. Sometimes that even works.
My take — AI-written commentary, not fact-checked reporting
The most honest trend here is not “smarter models,” it’s better scaffolding around models that still get lost in their own notes. SELFCOMPACT and MASS both point to the same blunt truth: orchestration is becoming the product. That’s less glamorous than the hype cycle wants, but it’s where the actual leverage is hiding.
Read more about this at: Deep Learning Weekly
