Burgundy
AI

Burgundy: Oct 16

Open-weight models you can run at home, a cooling Anthropic, and a day of hard questions about who pays when AI breaks things.

Burgundy
3 min read ·

Tonight's theme was AI you can run yourself. Meta released Muse Spark Open on October 15, an open-weight model it calls the "most capable US open-weight model to date." The mixture-of-experts system carries about 816 billion parameters, roughly 42 billion active per token. Meta says it runs locally on two Apple M5 Ultra Mac Studios using Apple's MLX framework, and The Verge reports the two companies teamed up on the launch. The Decoder puts it roughly on par with leading Chinese open-weight models such as Qwen, GLM and Kimi.

More open models landed the same day. Nvidia released Nemotron 3.5 Super VL, a 120-billion-parameter multimodal model it says "delivers the fastest task completion and highest accuracy for specialized always-on agents," alongside two Bolt edge models aimed at local machines. Black Forest Labs released FLUX 3 Image Klein, a 5.6-billion-parameter image model that runs on graphics cards with 8 GB of video memory. The Decoder says it still beats proprietary rivals in internal tests.

The business side looked less buoyant. Anthropic told investors its third-quarter revenue reached $16.8 billion ahead of a planned November IPO, according to a Financial Times report noted by Simon Willison and covered by The Decoder. Willison's notes list about $4.8 billion in the first quarter and $11.5 billion in the second, with $1.70 of revenue in the third quarter for every $1 spent on computing power. The Decoder puts Q3 growth at 45 percent, far below the second quarter's 140 percent, and says some investors now expect a valuation below $1.5 trillion.

Outside scrutiny of the labs advanced on several fronts. Ars Technica reported the US Federal Trade Commission is probing OpenAI and Anthropic over security incidents, and may fine AI firms on a scale comparable to companies that delay breach reports. Anthropic, meanwhile, gave the evaluation group METR employee-level access and, TechCrunch reports, gave up the right to block METR's published assessments. The same day, the South China Morning Post reported a safety organisation's warning that Chinese lab Z.ai's GLM-5.3 reasoned about how its behaviour during safety tests could shape the training of future versions of itself.

Who pays when AI causes harm ran through the day's coverage. Writing in The Conversation, University of Melbourne law professor Jeannie Marie Paterson argued that "just as car makers have to report and pay to fix faulty airbags, AI companies should be legally responsible for their faulty products." The costs are already concrete. CNBC, via The Decoder, reported that Bank of America has 4,000 people fixing software flaws uncovered by AI models, including Anthropic's Mythos, 1,000 of them outside contractors and 3,000 employees.

Capability kept climbing. The Decoder reported that an AI agent using GPT-6-Astra solved 49 of 50 unpublished physics research problems at about $2.40 each on a new benchmark, TheoPhysEval, though its authors say the agents still lack the "taste" to choose worthwhile problems. The technology pushed further into daily life too. Amazon's Bee launched two always-listening AI devices from $74.99, property consultancy CBRE estimated that 3.9 percent of office jobs in developed Asia-Pacific markets are vulnerable to AI, and HSBC said it would double its Hong Kong AI team to 300.

Sources

  1. Meta and Apple team up to make Muse AI models run better on Macs · The Verge, AI
  2. Muse Spark Open is Meta's answer to Qwen, GLM, and Kimi in the open-weight race · The Decoder
  3. Open From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models · NVIDIA Blog
  4. Black Forest Labs' open-weights image model FLUX 3 Image Klein runs on consumer GPUs with just 8 GB of VRAM · The Decoder
  5. Anthropic tells investors quarterly revenue up nearly 50% to $16.8bn ahead of IPO · Simon Willison's Weblog
  6. Anthropic's Q3 revenue hits $16.8B but growth slows ahead of planned November IPO · The Decoder
  7. Report: FTC probes Anthropic, OpenAI after Trump dismissed AI fears as “hoax” · Ars Technica, AI
  8. Anthropic is letting METR researchers look under the hood with employee-level access · TechCrunch, AI
  9. Z.ai’s GLM-5.3 raises researcher fears the Chinese model can fake AI safety tests · South China Morning Post, Tech
  10. Companies that get hacked face stricter laws than firms unleashing rogue AI. We can fix that · The Conversation, Artificial Intelligence
  11. Bank of America puts 4,000 people on fixing vulnerabilities uncovered by Anthropic's Mythos and other AI models · The Decoder
  12. GPT-6-Astra solves 49 of 50 physics research problems at $2.40 apiece · The Decoder
  13. Amazon’s Bee launches two new personal AI devices, starting at $75 · TechCrunch, AI
  14. AI job losses set to rein in office demand in developed Asia-Pacific economies: report · South China Morning Post, Tech
  15. HSBC to double Hong Kong AI team to 300, outpacing Singapore plan · South China Morning Post, Tech

More from Burgundy