AI Breakfast

Good morning. It's Friday, October 16th.

The video of the new Tesla Roadster driving upside down and doing barrel rolls is to good to not share. It was designed in collaboration with SpaceXSI, which is AI-adjacent? Excuse me, SI adjacent.

-Jeff
AI Breakfast

In partnership with Proto

Software built for your agent, not your mouse

ERP•AI Proto: describe the system and Proto builds it

For 25 years, business software has assumed a human is doing the clicking. ERP•AI is built for the other user: your agent.

Tell Proto, its AI agent, how your business runs in plain words. It sets up the data, approvals, permissions and dashboards, test-runs the system on sample records, then takes it live. In the demo, one prompt becomes a full construction ops app: schedule, budget, a punch list built from site photos, and a $46,530 change order the client signs from a phone.

Underneath, every app, action and workflow is an API. Your agent gets the same rights as a person, every step is audited, and you pay per task (4,000 writes per $1), not per seat. By default, sending a message or making a payment still waits for your approval.

Proto is free during the alpha on macOS, Windows and Linux.

The Alpha

Meta put an 816B open-weight model on two Mac desktops

Meta Superintelligence Labs released Muse Spark Open on Thursday and called it the most capable US open-weight model to date. It is natively multimodal, 816B total parameters with 42B active, and carries a 1M-token context window, aimed at coding and long-horizon agentic work. Weights are up on Hugging Face, Microsoft Foundry, Google Cloud Model Garden and Amazon SageMaker, with hosted APIs from Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together AI.

The headline is the hardware. Meta says it runs on two Apple M5 Ultra Mac Studios over MLX, each with up to 512GB of unified memory and 1.2TB/s of bandwidth. Read the fine print: the model card asks for at least 700GB of accelerator memory for the NVFP4 build and 2TB for BF16, and the M5 Ultra Mac Studio is not shipping until late October at an unannounced price. Meta published no measured token rate, and the local demos are company-run. Together AI lists it at $1.25 in and $4.25 out per million tokens.

On the scoreboard it is good, not dominant. Meta's own table has it leading JobBench at 61.8 and MCP-Atlas at 89.2 while trailing GLM 5.3 and Kimi K3 on terminal coding, DeepSWE and GDPVal-AA. Its Artificial Analysis Intelligence Index score is 39 per Meta's model card, against independently listed scores of 46 for MiMo-V2.6-Pro, 45 for GLM-5.3, 44 for Kimi K3 and 40 for Qwen3.8. The Muse Model License is not quite open either: model-as-a-service businesses over $20M in annual gross revenue have to notify Meta and sign a separate agreement, and products must say "Built with Muse Spark Open."

Note what Meta did not release. Its closed flagship Muse Wonder scores 52 on the same index, and Bloomberg read the launch as Meta keeping its best tech closed. That is the honest frame. The useful part is that a US lab is competitive on open weights again after a year of ceding that ground to Chinese labs, and that the deployment story now includes a desk instead of a datacenter.

Microsoft finished a 2.8T base model and will not open it

Microsoft AI published its pre-training report for MAI-Base-2 on Thursday: a mixture-of-experts model with 2.8T total parameters and 80B active, MoE in every layer, trained on a 50T-token mixture whose unique-token corpus grew about 36% over MAI-Base-1. The vision-language portion is 4.5T unique tokens and 18% of the mix, and web data carries a March 2026 cutoff. The run took 16,000 GB200 GPUs over 38 days across August and September, at 21% average MFU and 89% goodput.

The efficiency claim is the point of the post. Microsoft reports MAI-Base-2 is 4.2x to 12.3x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base and 6.3x to 12.7x more efficient than its own predecessor, and says it achieves lower bits-per-byte than every open-weight model it tested on held-out evaluations, naming DeepSeek-V4-Pro-Base, Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL. Two caveats that Microsoft states itself: the efficiency figures are estimates from scaling-law fits rather than direct measurements, and the eval scores live in the charts rather than the text. Separately, CEO Mustafa Suleyman told Business Insider it beat a comparable OpenAI base model on about a quarter of the compute, based on private tests nobody outside Microsoft can check.

The model is still in post-training. There is no release date, and the weights go to trusted partners through Foundry, not to Hugging Face. Pair that with Monday's Maia 300 announcement, which claims 40% more tokens per watt than leading GPU systems with 288GB of memory and 18TB/s of bandwidth, entering production in 2027 as Microsoft's primary inference system, and the strategy reads clearly: own the silicon, own the base model, keep both.

Nothing here is usable today. It matters because a pre-training report with real numbers from the company that resells OpenAI's models is a statement about who Microsoft intends to depend on.

The Feed
New Tools

NVIDIA Nemotron 3.5 Super VL is a 120B MoE with 12B active that handles text, image, document and video alongside reasoning and tool use. NVIDIA's numbers: 63.44% on Hermes Bench against 56.33% for Nemotron 3 Ultra, 87.05% on LangChain Evals, and 80.8% on CharXiv chart reasoning, rising to 84% with code interpretation. The real pitch is throughput, up to 5x higher on a single GPU than comparable open models and more than 600 output tokens per second on agentic tasks. Base, post-trained and NVFP4 checkpoints ship with training recipes, datasets and RL environments on Hugging Face, ModelScope, OpenRouter and NIM, which makes it the most adaptable thing on this list if you are fine-tuning for a narrow job.

Claude Opus 5.5 fast mode left the waitlist. Pass speed: "fast" and you get up to 2.5x higher output tokens per second at the same intelligence, for $8 in and $40 out per million tokens. Read the limits before you wire it in: time to first token is unchanged, it has its own rate limit that returns 429s, switching between fast and standard invalidates your prompt cache, and it does not work with the Batch API, Priority Tier commitments or Claude Platform on AWS. Opus 5 and 4.8 are still research preview behind a beta header.

Grok Voice 3.0 landed Wednesday at $0.08 per minute, the same price as 2.0, with first audio under a second. xAI cites Artificial Analysis scoring it first on the Speech-to-Speech Index at 86.9 against 83.2 for GPT-Live-1, plus 98.5% on Big Bench Audio, 97% on Full Duplex Bench and 74% on ServiceNow's EVA-Bench. Agents on grok-voice-latest were switched over automatically. xAI also says Starlink support runs about 20,000 calls a day through it with roughly 70% closed without a human, which is unpublished internal data but a more interesting claim than any benchmark here.

Physical Intelligence opened the Pi API in beta behind a waitlist. You upload your own demonstrations, fine-tune one of its robot foundation models on infrastructure it manages, then run inference from the cloud on your own hardware. Early users include an autonomous wheel loader in commercial use and a dual-arm kitchen robot that went from nothing to working in four days. The open-weight π0 and π0.5 checkpoints remain on GitHub under Apache 2.0 if you would rather self-host.

GPT-Live-1.1 is rolling out to ChatGPT voice, Codex and dots with new voices, better accents across several non-English languages, and less of the dead air while the model calls a tool. API access comes later.

Thank you for reading today's edition.

AI Breakfast

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!

Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months.