Deep Learning Weekly: Issue 468
Deep Learning Weekly Miko Planas
Deep Learning Weekly issue 468 is out, with Grok 4.6, TutorMoments, and new papers. The sharp bit: AI tutors still over-help, and coding agents are nowhere near done.
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
Deep Learning Weekly’s 468th issue is a grab bag, but the center of gravity is clear: agentic AI keeps getting pushed into harder, messier jobs, and the tests are finally starting to bite back. This week’s lineup includes xAI’s Grok 4.6, a local Meta model called Muse Glimmer, OpenAI’s expanded Daybreak cyber tiers, and a batch of work on tutoring, search, and code refactoring.
The most interesting practical story is probably Grok 4.6. xAI says it tuned the model for long-running agents and interactive builds, and that it matches GPT-5.6 Sol at 61 on the AA Intelligence Index. It also leads on GDPval-AA at 1753 and AA-Briefcase at 1577, while keeping pricing unchanged at $2 and $6. That’s the sort of release note that matters to people building with these systems, not just benchmarking them.
Meta’s Muse Glimmer is the opposite move: smaller, local, and always on. The company says the 30B agentic model was distilled from Muse Spark, quantized to under 20GB, and made fast with speculative decoding that can reach up to 3.1x faster decode. Meanwhile, Google DeepMind is pushing sign-language AI into consumer tools with SL2T, which was trained on more than 100,000 hours across over 50 sign languages and now powers ASL dictation in Gboard and Live Transcribe on Pixel 11.
The papers section is where the uncomfortable truth shows up. Ai2’s TutorMoments, built from 462 real math tutoring transcripts and more than 1,500 teacher-annotated decision points, finds that LLM tutors systematically over-help instead of nudging students into productive struggle. And in coding, SWE-Bench ProMax raises the bar with 170 multilingual refactoring tasks across seven languages, only to see the best frontier model resolve just 41.2% of them. That’s not saturation. That’s a warning label.
On the security side, OpenAI is widening Daybreak into Blue and Red access tiers and rolling out GPT-5.6-Cyber, which completes 95% of advanced offensive-security requests compared with 1.5% for the guarded base model. Anthropic is also preparing to watermark Claude-generated text and files to satisfy the EU AI Act’s Transparency Code. The week’s throughline is hard to miss: the industry keeps widening AI’s reach, but the evaluation problem is getting more honest at the same time.
My take — AI-written commentary, not fact-checked reporting
The real story here isn’t that models are getting smarter; it’s that people are finally testing them on jobs that resemble work. Tutoring, refactoring, cyber, search — all the places where hallucination turns expensive fast. The age of “good enough” demos is ending, and the bill is arriving with interest.
Read more about this at: Deep Learning Weekly