- A new benchmark called GDPx scored frontier models across 11 expert domains. The top result was 12%, and safety filters bent the rankings in ways the authors had to work around.
- DeepSeek open-sourced a coding super-agent prototype: one persistent router handing work to many parallel workers.
- OpenAI will put ChatGPT and Codex on a shared usage allowance starting in December, so your coding agent and your chat habit will compete for the same budget.
- Apple Intelligence is finally reaching China, with developer testing on November 4 and Alibaba's Qwen behind selected features.
- A security vendor says AI-assisted research turned up 28 previously unknown flaws in robotics software in under two days, and most of them were in the tooling, not the robots.
- For a map of who sells what in this market, start with the AI automation tool landscape.
The weekend and Monday had one thing in common, and it was not a model launch. It was measurement. A new benchmark tried to score models on work that people get paid for. A research team claimed a training method that reaches the same result with a fraction of the data. A security firm pointed a model at robotics code and counted the holes. Even the commercial news was about counting, with OpenAI moving to a single usage meter across two products.
When an industry starts arguing about scoreboards, it usually means the demos stopped settling arguments.
The Front Page: a benchmark for work that pays, and nobody is close
SemiAnalysis, the chip and infrastructure research firm, launched GDPx on November 2, a benchmark built to measure general agentic capability across 11 domains the firm picked for their economic value: chip design, cybersecurity, biology, legal work, quantitative trading, materials science, software engineering and more. The framing is deliberate. Most benchmarks test exam questions. This one tries to test jobs.
By SemiAnalysis's own scoring, Anthropic's Opus 5.5 led overall with 12%, winning quant trading, hardware design and chip design. The firm says GPT-6 Astra led on recursive self-improvement and materials science, Gemini 4 Carbon topped biology and software engineering, and models from OpenAI and Anthropic took most of the upper ranks. One detail is more interesting than the leaderboard: the report attributes some low scores to safety classifiers refusing to engage, and says that excluding those domains reshuffles the ranking. In other words, part of what this benchmark measures is how willing a model is to try.
Hold the number loosely. These are one firm's tasks, one firm's rubric, and no independent reproduction yet, from a company that sells research to the same industry it is grading. A 12% headline is also commercially convenient for everyone: it tells investors there is enormous room left to build, and it tells safety advocates the systems are not ready. Both camps can quote it happily.
What it means: The gap between "scores well on benchmarks" and "can be trusted with a domain" is now something you can point at in a budget meeting. That is useful. But do not read 12% as a verdict on the automation you were planning, because almost nothing in GDPx resembles reconciling invoices or drafting a customer reply. The practical lesson is the one benchmark builders keep relearning: the only evaluation that predicts your results is the one made of your tasks. Write down ten real jobs, grade the model on those, rerun it every time you change the prompt or the model.
Releases & Features
DeepSeek open-sourced a super-agent prototype. The dsh-superagent-prototype is orchestration code: a persistent router that reads requests against a GitHub repository and farms the work out to parallel worker sessions. The team reports that on simulated tasks, the slowest 5% finished in 15.7 minutes against 29.5 minutes for a single-session baseline. Its maintainers are clear that this is research code rather than a supported product, and that it acts only on explicit requests or rules the team writes down.
OpenAI is merging ChatGPT and Codex usage. Per its updated usage documentation, eligible work in both products will draw on one shared allowance from December. Ordinary chats are not supposed to count against it, and Plus, Pro and Business users will see remaining allowance and reset times in Settings.
Apple Intelligence reaches China. Apple confirmed the first features for Chinese users, with developer testing from November 4 and a December rollout alongside iOS 27.3. Alibaba's Qwen supports selected features. The redesigned Siri announced at WWDC is not in this phase.
What it means: The DeepSeek release is the one to read if you build anything agentic, because router-plus-workers is becoming the default shape for long tasks, and here it is in public code you can study rather than a diagram in a keynote. The OpenAI change is the kind of pricing detail that quietly reorders how teams work: a coding agent grinding through a large repository can eat an allowance a colleague was counting on. If your team shares a plan, find out in December whose job gets throttled first.
In the Lab
Randall Balestriero and Yann LeCun published a paper on JEPA, short for joint embedding predictive architecture, asking when separately trained text and image models end up with compatible internal representations. Plain English version: instead of training a model to predict the next word or the next pixel, you train it to predict useful summary features of a related view of the same thing, which lets it skip surface detail it never needed.
The authors prove alignment holds under stated assumptions about how those views are chosen, and report that on the same dataset their "good views" approach needed 42 times fewer samples than next-token prediction to reach comparable alignment. They also report that a text JEPA with 109 million parameters beat the language models they tested, up to 72 billion parameters, on their alignment measurements.
What it means: Those are alignment-metric results on specific datasets, not evidence that small models beat big ones at the work you care about, and nobody should swap a production model on the strength of one paper. The direction is what matters. Every result that cuts the data bill makes capable models cheaper to train outside the handful of companies that can afford a billion-dollar run, which is the quiet mechanism behind the open-weight releases that keep surprising people.
The Oversight Desk
As of November 1, US school districts can ask for enforceable AI privacy terms inside their Microsoft contracts, under a standard the company announced with the American Federation of Teachers and the United Federation of Teachers. It limits the use of student and educator data for AI training and advertising, adds transparency duties, and requires human oversight on consequential decisions. The Associated Press has the additional safeguards.
The catch is in the verb. Districts can request the standard in new or existing agreements, which means it does not apply the same way everywhere by default, and the districts with the least legal capacity are the least likely to ask. A protection you have to know about and negotiate for is a weaker protection than a floor.
What it means: Contract language is turning into the real regulation of AI data use while legislatures argue. That is worth copying outside education. If you hand a vendor data you do not own, whether that is customer records, patient details or employee files, the question is not whether the vendor has a trust page. It is whether your contract says the data cannot be used for training, who gets told when something goes wrong, and which decisions require a person. Ask for it in writing, because the terms exist now and they tend to go to whoever asks.
If the whole industry spent the weekend arguing about scoreboards, build yourself a small one. Tell BYOBot what you want an agent to do and get back a test set you can grade it against.
On the Radar
Smaller moves worth a glance, with the sources if you want to go deeper.
- 28 robotics zero-days in under two days. Security vendor VicOne says AI-assisted research found that many previously unknown flaws across the physical-AI stack, with about three quarters in development, training or simulation tooling rather than deployed robots, including one that stopped a remote-stop command from working. Read it as a vendor's sponsored report, not an audit. The Robot Report.
- One robots.txt signal for training preferences. OpenAI says it will support a shared signal so site owners can express one AI-training choice to multiple providers instead of maintaining a separate rule per crawler. Useful only if rivals adopt it. OpenAI.
- The xAI data-center pollution suit proceeds. A federal judge let an NAACP case over allegedly unpermitted gas turbines at the Colossus facility move forward. The ruling was procedural and settles no question of liability. TechCrunch.
- A haiku written in orbit. Google ran its Gemma 4 model on a data-center-class TPU aboard a prototype satellite. It is a real milestone and nowhere near an orbital data center, with radiation, heat and servicing all unsolved. SpaceNews.
- The small one: cancer screening in Panama. Korean firm Noul is supplying miLab CER, a compact device that runs cervical cell slide staining, imaging and AI analysis in about 20 minutes, to Panama's national public health system, aimed at facilities with no specialist pathologist on site. The company reports 93.9% sensitivity and 97.8% specificity on its own evaluations. The Herald Business, with Noul's product page for the specifications.
The Bottom Line
Four stories, one habit. GDPx counts what models can finish, the JEPA paper counts what training costs, VicOne counts what breaks, and Microsoft's school terms count what a vendor is allowed to keep. None of them is a launch, and together they say more about where this is heading than any launch did. The uncomfortable read is that the industry is still inventing its own yardsticks, which is why your own small, boring test set is worth more than a leaderboard. Pick one job, measure it honestly, then automate it. The workflow directory has plenty of candidates that pay for themselves inside a week.
Frequently Asked Questions
-
No, it means the test is hard and the scope is wide. GDPx asks models to carry out expert work across 11 domains including chip design, cybersecurity, biology, legal and quantitative trading, and SemiAnalysis scores partial credit against its own rubric. A low score on frontier expert work tells you very little about whether a model can draft your weekly report or reconcile an invoice list, which are the tasks most people are automating. Treat it as a ceiling check on autonomous expert judgment, not a verdict on everyday automation, and keep your own small evaluation set for the jobs you run. That difference between impressive and useful is the whole argument in from generative to functional AI.
-
JEPA stands for joint embedding predictive architecture. Instead of training a model to predict the next word or the next pixel, it trains the model to predict useful summary features of a related view of the same thing, which skips a lot of surface detail the model never needed. Sample efficiency is how much data a model needs to reach a given level of performance, and it matters because data and compute are the two bills everyone pays. A method that reaches the same result with far fewer examples lowers the cost floor for smaller teams and open models, not only for labs that can fund enormous training runs. If the vocabulary is new, our plain-English guide to prompt engineering is a gentler place to start.
-
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.
