Deep Learning Weekly: Issue 476
Deep Learning Weekly Miko Planas
Gemini 4 Argon beat Claude Opus 5.5 on a coding benchmark and shipped to 650-plus defenders. Deep learning’s newest flex is agents, evals, and a lot of token math.
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 issue is really a snapshot of where the field is pulling hardest: frontier models, agent tooling, and the messy business of making systems behave. The headline belongs to Gemini 4 Argon, which scored 77.9% on DeepSWE v1.1, ahead of Claude Opus 5.5’s 74.2%, and went first to 650-plus Fairwind Program defenders at 2/10 per million tokens.
There’s also a clear pattern in the enterprise releases. Cloudflare’s Clef adds Apache-2.0 decision models with a vision encoder and a 64k context window, and it’s pitched as faster at classifying a domain than gpt-oss-120b. Cohere rebuilt North around a new agent harness, while AWS opened up Strands Decider 2B, a laptop-runnable model that returns confidence-scored choices instead of free-form text. Different products, same message: don’t make every model do everything.
The research side leans even harder into that theme. One paper argues that cross-tokenizer distillation works better when supervision is kept tight and reliable, not bloated with extra alignment. Another finds agentic retrieval can beat standard dense retrieval by 8.7 nDCG@10 points, but at a steep price: 107.4 seconds on average, plus 764.1K input tokens and 5.8K output tokens per query.
And then there’s the ecosystem scaffolding around all this. Google is pushing EmbeddingGemma beyond text into images, audio, and video. Ai2 shipped Olmo-core 3 for trillion-parameter mixture-of-experts training. On the ops side, Google’s long-term memory recipe pairs Valkey with AlloyDB AI and claims token spend can drop by up to 70%. The week’s through-line is simple: capability is still rising, but the bill, the plumbing, and the evals are getting louder too.
My take — AI-written commentary, not fact-checked reporting
The industry still loves pretending bigger and broader is automatically better, and the week’s papers quietly disagree. Tight supervision, explicit decision models, and ugly little cost controls are doing more useful work than another round of “agentic” branding. The adults in the room are counting tokens, not vibes.
Read more about this at: Deep Learning Weekly
