Deep Learning Weekly: Issue 462
Deep Learning Weekly Miko Planas
OpenAI and Anthropic both pushed hardened, government-vetted models this week, while smaller labs chased speed and cost instead of scale. The split says a lot about where AI is actually headed next.
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
Security is suddenly the flex. OpenAI's limited preview of GPT-5.6 Sol, Terra, and Luna comes wrapped in government coordination, a layered safeguard stack, and 700,000 GPU hours of automated red-teaming — numbers meant to signal that this is the most locked-down model the company has shipped. Anthropic answered in kind, restoring Claude Fable 5 and Mythos 5 after US export controls lifted, but not without adding a beefed-up cybersecurity classifier and a new cross-industry jailbreak severity framework. Both companies are quietly building the compliance infrastructure that regulators, and probably lawyers, have been asking for.
Anthropic didn't stop there. Claude Sonnet 5 arrived as the most agentic Sonnet yet, closing the gap with Opus 4.8 on coding and tool use enough that the internal tier system is starting to blur. And Claude Science, a research workbench stitching together PubMed, Jupyter, and cluster compute with more than 60 scientific skills, reads like Anthropic's bet that scientists will trust an AI more if its outputs are auditable and reproducible rather than just fast.
Meanwhile, the small-model crowd kept proving that bigger isn't always the point. Liquid AI's LFM2.5-230M is its smallest open-weight release yet, and it still hits 213 tokens per second on a Galaxy S25 Ultra while beating larger models on tool use and data extraction. Google leaned the same direction with Nano Banana 2 Lite, generating images in four seconds for $0.034 per thousand, alongside Gemini Omni Flash for developer video work at ten cents a second. Frontier labs are chasing safety certifications; everyone else is chasing latency and price per call.
On the research side, a widely shared explainer on scaling laws walks through why the tidy power-law curves from Kaplan to Chinchilla get messy once data constraints enter the picture — fitting them cleanly in practice is harder than the charts suggest. A new paper called MRAgent challenges the standard retrieve-then-reason memory pipeline for LLM agents, modeling memory as a graph of cues, tags, and content that gets actively reconstructed during reasoning rather than pulled up wholesale; on LoCoMo and LongMemEval it claims up to 23% gains with lower token and runtime costs. And Unlimited OCR tackles a less glamorous but real problem: end-to-end OCR models like DeepSeek OCR slow down and balloon in memory as output gets longer, so the authors swapped in a sliding-window attention mechanism that keeps the KV cache constant, letting it transcribe dozens of pages in a single 32K-token pass.
My take — AI-written commentary, not fact-checked reporting
The pattern here is obvious once you stop staring at benchmark charts: the labs with the most compute are spending it on red-teaming and export compliance, not raw scale, which tells you regulators have already won more ground than the hype cycle admits. I'd rather see that 700,000 GPU hours spent proving a smaller open-weight model like LFM2.5 can do the same job on a phone, because that's the version of AI progress that actually reaches people outside San Francisco and doesn't need a government coordination call to ship.
Read more about this at: Deep Learning Weekly