Key points
- OpenAI's flagship models — Google would of course rather everyone used its Gemini, with Google Cloud services behind it.
- DeepSeek's Flash — DeepSeek put v4.1 Flash 1017 on OpenRouter, positioned as a speed-optimised lightweight version.
- Four Mac Studios — The most interesting part is the pair of numbers on power and memory.
- A model that only echoes — Simon Willison released llm-echo 0.5, a small version bump to his debugging plugin for the LLM CLI.
This is the Daily AI Stew, brought to you by Robin on 咸话咸说.
OpenAI’s flagship models, now on Google’s shelf
Google would of course rather everyone used its Gemini, with Google Cloud services behind it. But OpenAI and Google have reached an arrangement: OpenAI’s frontier models will land inside Google Cloud’s services. For a big cloud vendor that has long pushed only its own models, putting a rival’s flagship on the shelf is the whole story — the choice moves from the vendor to the customer. Source: arstechnica.com
What the shelf actually looks like: over the next few weeks OpenAI’s latest models join Google’s Model Garden. The garden is, plainly put, one model catalogue in the cloud — customers do not have to shop around vendors, they pick one and call it. It already holds Google’s own Gemini family plus a batch of open models such as Mistral, Llama and Gemma, and Anthropic’s Claude has been on it for a while.
The part that touches daily work is the coding tooling: the GPT family will soon appear in Antigravity, Google’s agent-driven coding tool. When you write code, the model dropdown gains a GPT row. OpenAI stresses that Microsoft remains its primary cloud partner, and confirms the full current GPT 6 / 6.1 lineup — Astra, Sol and Luna — will all be available on Google Cloud. Astra naturally costs the most to run, but Google’s systems can move workloads dynamically between LLMs based on performance, latency and efficiency: the same job can hop between models, whichever is the better value at that moment, without locking yourself to one vendor.
DeepSeek’s Flash, aimed at coding efficiency
DeepSeek put v4.1 Flash 1017 on OpenRouter, positioned as a speed-optimised lightweight version. It launched with a headline score: it beats Claude Sonnet 5.6 on 10 of 11 coding benchmarks. Community feedback clusters on the same points — value for money, generation speed (tokens per second and time to first token), and a visible chain of thought. Source: OpenRouter
As an open-weights model, developers can read the reasoning directly. People already use it as their daily driver for agent-swarm experiments, CLI agents and one-shot iOS app generation, reporting efficient token use and a good fit for multi-day autonomous agent workflows. The vendor calls it the newest deepseek-flash model, replacing V4-Flash and V4-Flash-Vision-Exp and adding native multimodality; it is listed on OpenRouter, OrcaRouter and other aggregators, billed per token at a sample price of $0.15 per million input tokens and $0.60 per million output.
The benchmarks are contested, though: the official page claims 10/11 against Sonnet 5.6, while some users note it still trails Sonnet and Gemini on other suites, and argue the “feels better” impression comes mostly from speed and the inspectable CoT. Others say its speed and price close the intelligence gap with top-tier models like Opus.
Four Mac Studios, running a trillion parameters
Apple’s machine learning research team published an MLX-based distributed inference demo: a four-Mac-Studio cluster running Meta’s roughly one-trillion-parameter Muse Spark Open at over 36 tokens per second. Source: Apple Machine Learning Research
The most interesting part is the pair of numbers on power and memory. A single M3 Ultra Mac Studio peaks around 200 W and an M5 Ultra around 300 W — less than a typical kitchen toaster — and one machine can carry up to 512 GB of unified memory. MLX is Apple’s framework tuned for Apple Silicon; it can run distributed inference and training across Macs over high-speed links such as Thunderbolt, so a cluster breaks past the single-machine unified-memory ceiling and fits models near a trillion parameters. The setup starts at one machine and scales on demand.
The model being run is no small thing either: Muse Spark Open is Meta Superintelligence Labs’ newest open-weights release, roughly 816B total parameters with about 42B active per token, native multimodal input, a one-million-token context window, and a design target of long-horizon agent tasks. Community measurements also show a single 256 GB M5 Ultra running DeepSeek 4.1 Flash 0910 smoothly at about 3,200 tokens/s prefill and 45 tokens/s generation, while a 128 GB M5 Max handles GLM 5.3 Flash.
Scepticism arrived too: commenters value the energy efficiency and the distributed-inference angle but point out that comparing 4-bit weights against Qwen’s native BF16 is not the usual comparison, and some read the whole thing as cross-brand promotion between Apple and Meta — a 512 GB M5 Ultra is not exactly a first-choice configuration.
A model that only echoes
Simon Willison released llm-echo 0.5, a small version bump to his debugging plugin for the LLM CLI. llm-echo provides an “echo model” that never calls a real LLM — it just echoes back the prompt it received, letting developers exercise the whole call path at zero cost. Source: simonwillison.net
This release extends the simulated token accounting to cover tool calls, system prompts and reasoning tokens at the same time. In practice it offers a way to avoid API charges, e.g. uvx --with llm-echo llm -m echo 'hi there' -o thinking 1 --usage, which with thinking enabled prints something like “Token usage: 2 input, 24 output”. The update is tool-chain work in support of the upcoming major LLM release he is preparing; it mainly serves plugin and CLI maintainers and has no direct impact on ordinary users — but putting tool calls, system prompts and reasoning tokens into one accounting view is exactly what you want first when asking why the bill went up.
Two-thirds of US doctors, and a $15 billion valuation
In the US, nearly two-thirds of doctors already use an AI-powered search engine to help judge their patients’ cases. That has pushed a previously little-known startup to a $15 billion valuation and started an industry argument about the future of patient care. Source: arstechnica.com
The company is OpenEvidence, headquartered in Miami and founded in 2022 by serial entrepreneur Daniel Nadler. By its own account, many of America’s one million doctors processed more than 42 million prompts last month, using the chatbot to trawl medical journals and clinical guidelines to help diagnose and treat patients. It grew by word of mouth and has seeped into an industry where few tools gain similar momentum. “We’re the first consumer internet app in healthcare that has really gone viral,” Nadler told the Financial Times.
For a company founded only in 2022, $15 billion means capital markets already treat doctors’ daily use as a revenue base. And precisely because it sits on the clinical-decision chain, where to draw the line between ad monetisation and patient care is the real focus of the argument. For doctors, what is saved is the time spent reading journals; for the industry, what changes is who gives the answer — journals and guidelines used to; now there is a model layer in between.
Meta has to name who signed off
In the copyright litigation over AI training, Meta has spent weeks pushing responsibility for downloading torrents onto employees. Now a court has ordered it to name which executives approved the decision to pirate books from shadow libraries such as Anna’s Archive and LibGen. This is TorrentFreak’s report. Source: arstechnica.com
The discovery order touches two cases at once: a class action by book authors including Richard Kadrey and Sarah Silverman, and a suit brought by publisher Entrepreneur Media. Magistrate Judge Thomas Hixon largely sided with the plaintiffs, holding that Meta must clearly identify the key decision-makers. After Meta employees who personally downloaded torrents testified, the order could pull in a wider circle — supervisors, managers and executives whom employees say green-lit the piracy.
Meta had hoped to avoid spelling out the chain of command, especially the leadership involved in early AI training. Per a Meta court filing, its initial large-scale AI dataset work was managed “informally”, and by its account no one with authority appears to have made any decision. Meta insists the only “decision” was employees “technically” downloading torrents. The point of that framing is to move liability from the decision layer to the execution layer. What the court wants to see now is exactly that chain: who nodded, who approved, who knowingly let it happen. In cases like this, whether a download happened is usually an execution detail; whether anyone authorised it decides whether the company is liable.
Apple rebranded, and hand-craft got valuable again
Apple unveiled a new brand identity for Apple Music, moving wholesale toward the Liquid Glass visual language — a direction Apple has pushed for a couple of years, so Apple Music following it is no surprise. Source: arstechnica.com
What is worth discussing is the other side: in an era saturated with talk of AI generation, brand design polished by hand matters more, not less. The arstechnica author did not spare himself — he admits he is still on MacOS Sequoia, grumbling at ever-larger corner radii and readability problems; but he also knows it is a wave that will pass, and that once the design starts hurting his work, or more importantly his timely access to security updates, he will probably give in and upgrade. The turn comes when he stumbles on Apple’s behind-the-scenes video: when generative AI makes visual production cheap, the scarcity of hand-crafted work goes up.
A cheap end and a control end
Today’s items describe two ends of one thing. On the cloud side the trend is cheaper orchestration — Google puts OpenAI’s flagship in its model garden and lets systems move workloads between models by performance, latency and efficiency; DeepSeek pushes coding-model value down another notch at $0.15 per million tokens and higher speed. On the local side it is the opposite: trading energy efficiency for control — four Mac Studios and open weights put a model near a trillion parameters on a desk, at the cost of accepting 4-bit quantisation comparisons and the reality that the expense is in machine count, not electricity. And a small tool like llm-echo, which only echoes and never reasons, is the most practical link in that cheapening chain: no API charges, and the tool-call, system-prompt and reasoning-token books still balance.
Looking one layer further out, AI is already rewriting places that are not servers. The chat box a doctor uses to check journals sits behind a $15 billion valuation, with the argument landing on the boundary between ads and patient care; in Meta’s copyright case the court is not asking who moved their hands but who knowingly nodded; doctors, judges and the employees who downloaded torrents all end up answering “who gives the answer, who is responsible”. Which is why Apple’s hand-polished visual language for Apple Music is not merely an aesthetic topic — when generation becomes cheap, people pay a little more attention to “someone actually did this”.
🎧 This episode is also available as a podcast — listen to Daily AI Stew · 2026-10-22.