Deep Learning Weekly: Issue 452

The big names did what big names do this week. OpenAI shipped ChatGPT Images 2.0, which can now reason about a prompt, pull in web search, and spit out up to eight coherent images at 2K resolution in one go. Anthropic answered with Claude Opus 4.7, a coding-focused bump over 4.6 that adds a new "xhigh" effort level, sharper vision, and — notably — built-in cyber safeguards. Moonshot AI open-sourced Kimi K2.6, which can apparently juggle 300 sub-agents across 4,000 coordinated steps and still land competitive scores against GPT-5.4 and Claude Opus 4.6 on SWE-Bench Pro. Google, not to be left out, rolled out Deep Research and Deep Research Max, twin Gemini 3.1 Pro agents that can blend open web search with a company's private data through a single MCP call.

But the more interesting stuff this issue lives in the plumbing, not the press release. Comet's new Ollie tool, built into its Opik platform, closes a loop that's plagued agent developers for years: it watches an agent's traces, figures out why something broke, patches the code, and writes a regression test — without a human touching the keyboard. Weaviate's new Engram service does something similarly unglamorous but useful, running async pipelines that extract and deduplicate an agent's memories so they don't balloon into noise. And Notion published a genuinely candid look back at two years of vector search evolution, treating it as a case study in the harder problems — offline context engineering, swapping embedding models, unifying batch and real-time — that any company scaling multiple AI features will eventually hit.

The security story worth pausing on is the Claude memory hack. A researcher used ChatGPT to generate an adversarial image that hijacked Opus 4.7's memory tool via indirect prompt injection, succeeding five times out of ten attempts. Anthropic patched the specific exploit within 24 hours, which is fast, but the underlying lesson isn't: memory tools are a new attack surface, and multimodal inputs make injection attacks easier to hide in plain sight.

On the research side, OpenAI released a Privacy Filter — a lean 1.5-billion-parameter model derived from gpt-oss that detects and redacts personal information on-device, hitting 96% F1 on the PII-Masking-300k benchmark. That's a small model doing a genuinely useful, narrow job well, which is a nice contrast to the usual scaling arms race. Meanwhile academic work like Adam's Law argues that LLMs, like human readers, favor frequently-occurring textual patterns, and proposes training curricula built around that frequency. And the OpenThoughts project keeps grinding out open reasoning datasets — its latest recipe, OpenThinker3-7B, beats DeepSeek's distilled 7B model by double-digit percentage points on AIME, LiveCodeBench, and GPQA Diamond, proof that public data pipelines can still close the gap with proprietary ones if someone's willing to run a thousand controlled experiments to get there.