4 AI drops worth watching: September 26
Agentic CUDA Kernel Optimizer: an open-source agent that rewrites and benchmarks GPU kernels
On September 24, developer bertaye released an open-source agentic CUDA kernel optimizer on GitHub, built with LangGraph. Given a workload description, the agent generates a kernel implementation, compiles it through a C++ test harness using NVRTC and the CUDA Driver API, and checks correctness against a NumPy reference before measuring latency with CUDA events. It can call Nsight Compute to inspect profiler counters and pull NVIDIA documentation for optimization guidance, then feeds those results back into the next iteration. The default model is GPT-5-mini at medium reasoning effort, and each run is logged to a results directory with the fastest validated kernel, a timing heatmap, and full execution history. The project explicitly states it does not benchmark against cuBLAS or other vendor libraries, and correctness only holds for the supplied test cases.
For anyone who has spent a weekend hand-tuning a GEMM kernel, this turns kernel optimization into a loop instead of a guessing game: describe the workload, let the agent iterate, keep the fastest correct version.
The take. This is a narrow, honest tool, and the README says so directly rather than overselling it: it optimizes individual kernels against cases the developer supplies, nothing more. The interesting part is the profiling loop, letting the agent read Nsight counters and NVIDIA docs mid-run turns kernel tuning into a genuine iterative search instead of one-shot code generation. Watch whether someone benchmarks it against cuBLAS or a hand-tuned baseline; without that comparison, any “found a faster kernel” claim stays unverifiable. If a fork adds that baseline and the numbers hold, this becomes a template for agent-driven low-level performance work beyond CUDA.
Definitions:
- NVRTC (NVIDIA Runtime Compilation): a library that compiles CUDA C++ code at runtime instead of ahead of time.
- Nsight Compute: NVIDIA’s kernel-level GPU profiler.
- LangGraph: a framework for building multi-step agent workflows on top of an LLM.
Chess Post-Mortem Skills: a Claude Code skill that narrates your own chess game back to you
Developer brumar released a set of Claude Code skills that turn a single chess game into an annotated post-mortem, including a narrated video. The workflow takes a lichess game link plus an optional audio recording of the player thinking aloud during the game, transcribes that audio locally with whisper.cpp, and aligns each spoken thought to the move being considered using PGN clock times. Claude then drives Stockfish itself, asking the engine the questions a human would (“why not the move I played?”) until the explanation holds up, and produces an annotated PGN, an HTML analysis board, and a roughly six-minute narrated video with subtitles. A full run, from engine sweep through finished video, takes about an hour, and the repo ships chess-video and chess-play as separate skills for narration and for playing Claude directly over PGN.
It replaces clicking through engine eval-bar drops with something closer to a coach reviewing the game with you, using your own in-game reasoning as the starting point rather than a guessed one.
The take. The mechanism worth noting isn’t the chess angle, it’s using a player’s own recorded reasoning as ground truth for what the agent should confirm or refute, a pattern that generalizes to any domain where someone already narrates decisions live: trading, debugging, even meetings. The stated hour-long runtime and the built-in verifier pass to catch hallucinations are both disclosed limits rather than hidden costs, which is rarer than it should be in this category of project. If someone strips out the chess-specific parts and keeps the audio-to-decision-alignment core, this becomes a generic post-mortem skill for any recorded process, not just games.
Definitions:
- PGN (Portable Game Notation): the standard text format for recording chess moves and clock times.
- Stockfish: a free, strong open-source chess engine used to evaluate positions.
- whisper.cpp: an open-source speech-to-text transcription engine that runs locally.
Meta’s Muse: an outside investigation finds a hidden OpenAI-labeled model inside the assistant
Following up on an earlier filesystem dig into Meta’s Muse assistant, blogger Pete of mouse.dev reports finding a session log routed to a model labeled azure/muse-special, distinct from the roughly fifteen versions of Meta’s internal model, called Avocado, that reportedly handle nearly every other session. The muse-special transcript carries a gpt_responses_v1 signature with an encrypted payload prefixed gAAAAA and tool-call IDs in OpenAI’s formatting style, both different from the Avocado sessions. The broader model catalogue shipped inside Muse’s agent daemon reportedly also lists Claude Opus and Sonnet versions, GPT-5.5 and GPT-5.6 variants via OpenAI, Azure, and Codex, and Kimi K3, alongside working Anthropic request-handling code and a proxy kill switch in the runtime environment. The reporter concludes muse-special is likely an OpenAI model served through Azure, though the exact version isn’t confirmed, and finds no evidence Meta is distilling from OpenAI or Anthropic’s raw outputs, since that reasoning arrives encrypted.
For a product marketed as Meta’s own personal AI agent, this is a reminder that the model users actually talk to can be a routing decision made server-side and invisible to them.
The take. This is well-sourced circumstantial evidence, not a confirmed disclosure, and it should be read that way: filenames, signature formats, and ID conventions, not a leaked system prompt. If Meta confirms or denies the routing in the coming weeks, this becomes either a minor footnote or a real story about frontier labs quietly reselling access to each other’s models inside consumer products. Either way, the shipped multi-provider client code and the kill switch show Meta built the infrastructure to swap model vendors without telling anyone, which is the more durable finding regardless of what muse-special turns out to be.
Definitions:
- Avocado: Meta’s internal name for the primary model powering Muse.
- Azure: Microsoft’s cloud platform, here apparently used to serve a non-Meta model.
Microsoft: Copilot drops the personal-assistant pitch, goes all in on the workplace
Microsoft is merging its consumer and workplace Copilot products into a single assistant aimed at corporate customers, according to Bloomberg reporting on an internal preview shown to a few dozen business and technology leaders in Seattle this week. The rebuilt Copilot is the result of a six-month engineering effort, absorbs design elements from the consumer version, and drops the marketing framing that positioned Microsoft as a builder of personal AI for individual users. The new Copilot is set to roll out in the coming weeks. The move effectively cedes the personal chatbot market to OpenAI, Google, and Meta.
It’s a public admission that a company with the distribution of Windows and Office could not make a consumer chatbot stick against ChatGPT, Gemini, and now Meta’s own push, and is retreating to the market it already owns: the enterprise seat.
The take. The bottleneck Microsoft is addressing isn’t technical, it’s positioning: running a split consumer brand and enterprise brand diluted resources and confused the pitch while OpenAI, Google, and Meta each built dedicated personal-assistant products. Folding everything into a workplace-first Copilot is a bet that distribution through Microsoft 365 seats beats competing head-on for individual mindshare. Watch what happens to Copilot’s consumer app in the following months; if it quietly ships fewer updates than the enterprise line, that confirms this is a full retreat rather than a rebrand.
Each of these is really about who controls the layer underneath the product: which model answers a prompt, which engine checks an agent’s work, which brand a company bets its assistant on, and whether the builder using the tool can see any of it.
More drops at dropwatch.ai. Want them in your inbox? Subscribe below.