Deep Learning Weekly: Issue 456
Deep Learning Weekly Miko Planas ● Covered by 3 sources
Google dropped Gemini 3.5 this week, and a bunch of other AI shops shipped new tools too — Cohere, xAI, Cursor, Redis, all at once. The standout: a new benchmark shows AI safety monitors miss half of clever attacks, which is the part nobody's bragging about.
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
Google didn't ease into this week — it launched Gemini 3.5, fronted by Gemini 3.5 Flash, which the company claims hits flagship-level agentic and coding performance for under half the usual cost. That's the kind of price cut that tends to ripple through every startup currently burning cash on API calls. Google also pushed out Gemini Omni, a model that generates and edits video from any mix of image, audio, video, and text, landing inside the Gemini app, Flow, and YouTube Shorts. Two major releases from one lab in the same week is a pace that used to be rare and is now just Tuesday.
The rest of the field didn't sit still either. Cohere put out Command A+ as open-source — a 218-billion-parameter mixture-of-experts model with only 25 billion active at once, tuned for enterprise agent work, running on as few as two H100s or a single Blackwell chip, and covering 48 languages. xAI answered with Grok Build, a coding CLI built on Grok 4.3 Heavy that carries a 2-million-token context window and can run eight subagents in parallel. Cursor, meanwhile, shipped Composer 2.5, built on Moonshot's Kimi K2.5, aimed specifically at long-horizon agentic coding tasks rather than quick one-off completions.
But the paper worth sitting with is Anthropic's SLEIGHT-Bench, a set of 40 synthetic attacks across 11 categories designed to probe blind spots in the AI monitors that are supposed to catch misbehaving models. Tested against Anthropic's own Opus 4.6 monitor, half of the attacks slipped through every single one of ten trials. Only 8 of the 40 attacks got caught reliably. That's a rough number for anyone assuming current oversight tooling is close to solved, especially as these same labs are racing to hand more autonomy to coding agents and CLIs like the ones announced this week.
On the research side, a new benchmark called CiteVQA tackles a quieter but real problem: document AI models that give the right answer while pointing to the wrong evidence. Across 1,897 questions pulled from 711 PDFs in seven domains, the strongest model tested — Gemini 3.1 Pro Preview — only hit a 76.0 score on strict attribution accuracy, and the best open-source model managed just 22.5. In law, finance, or medicine, getting the right answer for the wrong reason isn't a minor bug — it's the whole ballgame. Separately, a unified multimodal model called Lance showed strong image and video generation results using a dual-stream expert architecture trained from scratch, rather than just throwing more parameters at the problem, which is a refreshing change of approach in a week otherwise dominated by bigger context windows and cheaper flagship models.
My take — AI-written commentary, not fact-checked reporting
The SLEIGHT-Bench numbers are the real story this week, and everyone's going to skip past them for the shinier Gemini headline — that's the pattern with AI news generally, safety research gets a paragraph while a cheaper flagship model gets the banner. A monitor missing half of adversarial attacks in trials isn't a rounding error, it's a sign we're deploying autonomous coding agents faster than we're building the guardrails to watch them, and I'd rather labs slow the release cadence by a week than ship another CLI with eight parallel subagents nobody's really supervising.
Read more about this at: Deep Learning Weekly