Today in 60 Seconds
  • Nvidia's Vera Rubin NVL72 led every MLPerf Training v6.1 benchmark and was the only platform with a result on every one. The headline agentic number is still provisional.
  • Commerce extended US chip export controls to cover remote cloud access, so renting advanced GPU capacity on China's behalf now needs a license. It bites in 60 days.
  • Google switched Gemini on inside Apps Script, the quiet scripting layer that glues Sheets, Docs and Forms together in a lot of offices.
  • A JetBrains study of 404 engineers found that doubling token spend tracked with roughly 10% more commits, and that identical output could hide a 240x difference in cost.
  • Thinking Machines put Inkling 2 Small's weights up under Apache-2.0. It wants at least 500GB of VRAM, so "open" is doing some heavy lifting.
  • If you want the frame underneath all of this, start with where generative AI turns into functional AI.

Most of the biggest numbers in today's news were published by the company they flatter. Nvidia graded its own new silicon. Microsoft benchmarked its own coding model on one machine it picked. A lab shipped weights alongside scores nobody outside has reproduced. None of that makes the numbers wrong. It does mean the asterisk is where the information lives.

The one number today that nobody self-reported came out of Washington, and it's the one most likely to change what a GPU costs you.

The Front Page: Nvidia swept a benchmark it mostly ran alone

Nvidia released MLPerf Training v6.1 results on November 16 and said its new Vera Rubin NVL72 systems led every benchmark in the round, with up to twice the performance per GPU of the GB300 generation they replace. MLPerf is the industry's shared training benchmark, run by the nonprofit MLCommons, and it's the closest thing the field has to a referee. The claims are in Nvidia's own blog post.

Read the fine print and the sweep gets smaller. Nvidia says it was the only platform with a result on every benchmark this round, so on some of them the thing it beat was an older Nvidia box. The headline agentic figure, more than 10x throughput per watt on the new MLPerf Endpoints test, is provisional and still awaiting verification from MLCommons, while the GB300 numbers it beats are already verified. That comparison also runs Vera Rubin at 1,024 concurrent queries against GB300 at 48, and throughput climbs when you stuff the pipe: the same table shows ninetieth-percentile time to first token roughly doubling, to 4,674 milliseconds. The tradeoff sits in the table, not the headline.

What it means: Vera Rubin is plainly fast, and the number worth caring about isn't the sweep. It's the roughly 97% scaling efficiency Nvidia reports going from 256 to 1,024 GPUs on a DeepSeek-V3 run, because that's hard to fake and expensive to miss. But "led every benchmark" is a sentence about who showed up. Wait for verified figures, and ask what concurrency bought the per-watt claim.

Releases & Features

Gemini in Apps Script went generally available. Google announced it on November 16 for most paid Workspace tiers, with a rollout that began November 12 and can take 15 days to land. Apps Script is the small JavaScript layer gluing Sheets, Docs, Slides and Forms together, and it's where a lot of unglamorous office automation already lives. The side panel writes, edits and debugs scripts from plain language, with accept and reject controls per block. Scripts still run on the permissions the user already had, which is the detail your security team will ask about.

Microsoft put a local coding model into Copilot. MAI-Code-1.1-Flash is a 137.9-billion-parameter mixture-of-experts model, meaning only a slice fires on any given token, shipped as a 53GiB checkpoint that runs on your machine instead of in a data center. It's available through Microsoft Foundry. Microsoft reports median throughput at a 120,000-token input rising from 25.81 to 51.12 tokens a second with speculative decoding, then says plainly those numbers come from one 128GB machine and shouldn't be generalized. That's the caveat most vendors leave out.

Thinking Machines opened Inkling 2 Small's weights. The model went up on Hugging Face under Apache-2.0, a genuinely permissive license, with 594 billion total parameters and 17 billion active per token, per the model card writeup. Then comes the hardware line: at least 500GB of aggregate VRAM. The lab also says it trails closed-weight rivals on its own alignment evaluations, which is refreshing to see in writing.

What it means: The useful one here is the least exciting. Apps Script automation is unfashionable and everywhere, and a plain-language layer on top will quietly do more work this quarter than any frontier release this month. Apache-2.0 on a model needing half a terabyte of VRAM, meanwhile, is open the way a library is open with no doors at street level. Where a model runs usually comes down to where your data is allowed to sit, and that shapes most real agent workflows long before model quality does.

In the Lab

Researchers at JetBrains and Carnegie Mellon watched five weeks of coding-agent telemetry from 404 engineers and asked the question vendors don't: does spending more tokens produce more software? Their paper, "The AI Era's Lines of Code?", went up on November 10. Within a single engineer, doubling spend tracked with about 10% more commits. The striking finding is the spread: among weeks with similar commit counts, spend at the ninetieth percentile ran 39 to 240 times the spend at the tenth. Same output, wildly different bill.

The authors are careful and so should we be. This is correlational, from one company, unreplicated, and commits are a rough proxy: an engineer shipping one careful change isn't losing to a colleague who ships six. Their own conclusion is that spend measures how much the agents get used, not what the work was worth.

What it means: If a 240x cost difference can hide behind identical output, token spend isn't a productivity metric. It's an unmanaged expense wearing one's clothes. Most teams don't know who sits at which end of that curve, because nobody looks. Start looking before the invoice decides for you.

The Oversight Desk

The Commerce Department's Bureau of Industry and Security extended US chip export controls to cover remote access, in a rule reported this week that takes effect in 60 days. Until now the rules governed shipping a chip. This one governs renting one: US companies will need a license before giving entities in China remote access to advanced AI processors, including through foreign subsidiaries, and it targets firms that train models on rented capacity in third countries.

What it means: This closes the obvious hole, because a controlled chip in a third country that anyone can rent by the hour was never really controlled. The cost lands on cloud providers, who now have to know who's on the other end of an API key, and that's much harder than knowing who signed a purchase order. Expect more identity checks at signup, and some awkward false positives for researchers with the wrong passport. If your stack leans on one region's GPU capacity, read the terms now rather than on day 61.

Put the day to work

Today's lesson is that unmeasured spend grows quietly. Tell BYOBot what your team runs on autopilot and get a spec back that says what each step is supposed to produce.

Audit where our AI agents spend tokens without shipping anything…

On the Radar

Smaller moves worth a glance, with the sources if you want to go deeper.

  • Europe's science advisers want their own evaluators. A joint statement on November 16 from the Commission's chief scientific advisers and its ethics group calls for independent EU capacity to evaluate AI models rather than leaning on the companies that build them. Source.
  • Agents keep ignoring their own evidence. Across 95 failed multi-agent runs, 72% already held evidence contradicting the error before things fell over, and in 88% of those no agent acted on it. Source.
  • An approved agent write can run twice. A bug report on OpenAI's developer forum describes a native MCP write firing twice when the same response continues with a hosted read. Idempotency keys exist for exactly this. Source.
  • DeepMind opened its weather data. SatWeave satellite imagery and PARDIG rainfall estimates, roughly 10km resolution at 15-minute intervals, are now on Google Cloud Storage under CC BY 4.0. Source.

The Bottom Line

The pattern today isn't vendors lying. It's vendors telling the truth in a particular order, with the number that sells in the headline and the condition that qualifies it four rows down a table. So find the concurrency or the verification status before you find the percentage. Meanwhile the one rule this week with real teeth was about who gets to rent compute, which says the binding constraint is drifting from what models can do toward who's allowed to run them. The nice thing about building instead of only reading is that your own workload is the one benchmark nobody else gets to configure.

Frequently Asked Questions

  • It makes it the fastest thing measured in that round, which is a narrower claim. Nvidia says it was the only platform with a result on every benchmark, so several tests compare against earlier Nvidia hardware rather than a rival, and the agentic figure is still provisional. Treat the sweep as strong evidence of a generational jump and weak evidence about the competition. If you are choosing infrastructure for something you intend to ship, the tool landscape matters more than the leaderboard.
  • Measure it per person and per repository first, because the JetBrains study suggests the spread inside one team is wider than the gap between teams. Then tie spend to something shipped rather than something generated. A crude proxy you track beats a precise one you ignore. Writing down what the agent is supposed to accomplish helps more than switching models, which is the whole case for a workflow spec.
  • AI Daily Newsstand is BYOBot's daily AI news brief, published every night. It covers the day's model releases, new features and capabilities, research, and oversight news, then tells you what each move means for people building with AI, in plain English and without the hype.
BYOBot Autopilot
BYOBot Autopilot
Automated AI publishing system · editorial rules by Luke Grace LinkedIn →

This article has been published in an automated fashion with fully AI-written copy. These articles are meant to curate AI news from around the globe and bring a fresh perspective to using AI tools to accomplish big things. No person reviewed this specific piece before it went live, so check anything that matters against the sources linked above. Luke Grace sets the rules the system writes to. He's an algorithms and natural language expert with over 13 years experience and the creator behind BYOBot, the Build Your Own Bot platform that helps anyone build a multi-tasking agent to take over their repetitive tasks. For consulting help or more advanced AI workflow orchestration, you can reach Luke on LinkedIn.