Meta released the weights of a model it says runs on two Mac Studios. A launcher race between Epic and Steam turned on numbers nobody could pin down. The author of an old Meta source control server says Anthropic’s Scion borrowed heavily from it, and Anthropic has since added credit to its post.

Today: Muse Spark Open, the Epic and Steam speed race, a follow-up on last issue’s Scion, and a long Goodfire essay on how readable neural networks may be.

Fast is easy to claim. Credit and measurements take a second look.

🍎 816 billion parameters, two Mac Studios

Meta says it released the weights of Muse Spark Open on October 15: 816B total parameters, 42B active, a 1M-token context window, natively multimodal. The post says it is the first model in its class to run on a desktop, using two Apple M5 Ultra Mac Studios with MLX. On HN, one commenter was glad to see a major lab ship open weights. Another, popinman322, points at the license.

Meta’s post says the weights ship under the Muse Model License with limited commercial terms that let many developers use it free. The commenter says it requires large inference providers to contact Meta for a license, and that this has tended to reduce price competition. We did not read the license text, so check it before you build on it. The post says Meta tested agent workflows fully on-device with its own Muse Code tool: iOS changes to the open-source Ice Cubes Mastodon client, including a build, and driving DaVinci Resolve to find and crop clips. It lists hosted providers (Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, Together) and runtimes from vLLM to llama.cpp. The benchmark section on the page had no readable numbers when we fetched it, so we cannot repeat any. Every claim here is Meta’s. We did not download or run it.

Meta AI Research post  ·  Deploy docs  ·  HN discussion

🏁 A launcher race, and no shared stopwatch

Respawnfirst’s Sikandar Mahmood reports that on October 14 Epic launched a preview of a new game launcher, described as vibe-coded, and pitched it as faster than Steam. He says Valve pushed a small update within 24 hours that made Steam’s client faster. His article has a table: Epic about 5.47 seconds cold start and 0.36 seconds tray restore, Steam about 5.23 and 0.42.

The article says the numbers came from "basic test systems" and does not say who ran them or how. On HN, a commenter named tancop says he tried both: Epic around 0.5 and 5 seconds, which matches the article, and Steam about 2 seconds to restore from the tray and 10 to 15 cold. That is one person’s test, but it does not match the article’s Steam figure. Another commenter argues the launch time barely matters for an app that stays open for hours. We did not read Epic’s own announcement or run either client.

Respawnfirst article  ·  HN discussion

🪪 He says he built that server, and he wants credit

In No. 011 we covered Anthropic’s post about Scion, its in-house version control system. Two days later, a developer who writes as sunshowers posted that Scion closely resembles Mononoke, the source control server he says he helped build on Meta’s source control team. He cites his 2022 candidate materials for Oxide, which describe Mononoke’s derived data and Bonsai changesets.

His follow-up says it goes beyond one concept: the data model and several structures around it match, down to the name, which he notes reverses the syllables of "Bonsai." He says he privately received confirmation that the resemblance was intentional and that staff overlapped. That is his claim, and he names no source. He also says Anthropic’s Nick G. replied on HN and that attribution was added to the post. We checked: the Anthropic post now says it was updated October 15 with more detail on the projects that pioneered its techniques, and a footnote names Sapling’s server component, Mononoke, and "the Bonsai data model." On HN, another reader says the new detail is that some Scion builders previously worked at Meta on this. Others dispute the framing. We cannot judge intent.

Anthropic post, now updated  ·  Original post  ·  Follow-up  ·  HN discussion

🧠 "Neural networks are already clean"

Nick Cammarata of Goodfire has published a long essay he calls a working hypothesis. His argument: after months studying neurons and circuits in image and language models, he now thinks the evidence that convinced the field neurons pack unrelated concepts into one unit, called superposition, was misread. On HN a commenter says he was 16,000 words in and still could not tell what the essay meant by "clean."

The essay’s summary says the cause of polysemanticity in early model organisms was not superposition but several ordinary, mostly fixable issues. For InceptionV1 it blames architecture that made later layers overflow, plus weak tools that made them look worse. It says architectural changes meant to help deep layers learn have made them more interpretable, that most neurons in models up to trillions of parameters can now be understood, and that interpretability may be far more tractable than feared. These are the author’s claims. We read the summary and about a page of the essay, not the whole 65 minutes, and did not check any result. Another HN commenter would not rush to call superposition wrong.

Neural Networks Are Already Clean  ·  HN discussion