COBOL scores 72%, a fifth of Anthropic's shares for the public good, and websites locking agents out

In the three days to October 20, 2026, seven items from the AI and tech accounts this channel follows were worth carrying. OpenAI published test numbers for COBOL, Fortran and Visual Basic, the languages underneath a lot of banking and government software. Anthropic put a slice of its own shares into a fund for the public good, and explained how the slice grows. Ethan Mollick picked up a report on the large websites now shutting AI agents out. The window runs three days because the editions of October 18 and October 19 were not published.

Source scope: the fixed set of public AI and tech accounts this channel monitors, while the personal X connection is not linked. One of those accounts, Benedict Evans, can no longer be read on X, so its posts are absent from this edition.

OpenAI's models get much better at COBOL, Fortran and Visual Basic OpenAI's research post, and its post on X · 6,615 likes and 845,000 views on the post at capture · October 19

OpenAI published two evaluations on October 19, both aimed at the languages its enterprise customers still run. MultiLangBench asks a model to build a working SQLite-compatible database engine from scratch in each of 20 programming languages. The second asks it to migrate a COBOL program to C++, Java, Kotlin, Python or Rust while keeping the behavior identical, judged on hidden tests 1.

At maximum reasoning effort, GPT-6 Astra passes 87.4% of MultiLangBench tests on average, against 26.8% for GPT-5.6 Sol. The per-language gains are wide: COBOL 10.9% to 72.1%, Fortran 21.4% to 83.5%, Visual Basic 27.5% to 89.4%. Astra clears 83% in 19 of the 20 languages, and COBOL stays the hardest 1.

OpenAI reports a cheaper option alongside Astra. GPT-6 Luna averages 40.3%, ahead of GPT-5.6 Sol, at an estimated $0.77 per sample against $3.48, which OpenAI puts at roughly 78% lower cost 1. On migration, GPT-6.1 Sol scores 86.6% at low reasoning effort where GPT-5.6 Sol scores 5.4% at the same setting, and at 61% of the API cost; pushed harder, GPT-6.1 Sol reaches 97.9% and nearly saturates the test.

The stakes are the old systems themselves. OpenAI cites IBM's figures that COBOL supports more than 40% of online banking systems, 80% of in-person credit card transactions and 95% of ATM transactions 2. These are OpenAI's own evaluations, and the company says a production migration still needs expert review and testing. The benchmark suite and its per-language results are on GitHub 3.

Anthropic commits roughly a fifth of its shares to a public benefit fund Anthropic's announcement, and its post on X · 21,170 likes and 7.24 million views on the post at capture · October 18

Anthropic will commit 5% of its shares, subject to stockholder approval, to a new entity called the Anthropic Public Benefit Fund. For every 50% increase in the company's value above a set baseline, the board may add another 2% 4.

The fund is a wholly owned LLC, picked over a charitable foundation so the company can move between forms of giving. None of the equity it holds goes to lobbying or political donations, in the US or elsewhere 4.

The pledges already made push the total higher. The co-founders promised 80% of their equity at founding, and about half of employees give through a matching program, so roughly 20% of Anthropic's shares end up dedicated to the public good 4. Anthropic notes that escalating equity awards are ordinary for tech executives, and says this equity goes to the public instead.

What the money is for is still broad. Anthropic names large donations to established nonprofits and direct distributions to particular populations as possibilities, with labor-market disruption as the example 4. The company has already committed $200 million to research and work on that problem 5. Governance, the valuation milestones and a public benefit report are promised for the coming months, which leaves the board's discretion as the operative rule for now.

Chollet: agents write the code, and the bottom line is where it was François Chollet's posts on X · 2,775 likes and 277,230 views on the post at capture · October 17

François Chollet, who created the Keras library and runs the ARC Prize, posted four observations about the "total" automation of software engineering. There are more software engineers than in 2025 and demand is growing. Engineers have never worked harder. Agents write close to 100% of the code. And the bottom line of the companies employing all of it has barely moved 6.

His explanation is systemic: automating one part of a system creates friction somewhere else and pushes the bottleneck along. The claim rests on his own view of the industry, and he offers no measurement with it.

He then applied the test to a claim he had seen about AI gutting SaaS companies. Chollet writes that he searched for data and found none to support it, and points to steady revenue trends at Adobe, Workday, Zoom and Salesforce alongside flat software-engineering headcount at those companies 7. He offers the four companies as examples rather than as a study.

Mollick: the web is closing its doors on the agents it was built for Ethan Mollick's posts on X, quoting a thread, and The Verge's report · 780 likes and 68,942 views on his post at capture · October 17–18

The Verge reported on October 17 that the sites agents were meant to shop and bank on have started blocking them. Yelp admits only agents that pay for its data-licensing program, Delta stops agents booking tickets on a user's behalf, JPMorgan Chase and Charles Schwab block agent access to accounts on security grounds, and Amazon has banned Perplexity's shopping agent and Meta's Muse 8.

Ethan Mollick, a Wharton professor who studies how people use AI, quoted a thread listing those blocks the next day and drew a conclusion from it. A strategy of shutting agents out looks fragile, he wrote, in the era of AIs that can use your computer on your behalf 910.

His follow-up draws the line he would keep. Handing today's agents control of a bank account is a bad idea given how little is known about agentic security and prompt injection, he writes, and finding a secure way to work with agents will serve everyone better than a ban 11.

The legal ground under the blocking is unsettled. A judge ordered Perplexity to stop its agents shopping on Amazon, an appeals court overturned that in August 2026, and Amazon has since filed an amended complaint 8.

Graham: publishing behind a paywall now means being invisible Paul Graham's post on X · 3,038 likes and 203,908 views on the post at capture · October 18

Paul Graham, the essayist and Y Combinator co-founder, put the claim in two sentences. "To publish something on a paywalled site is to make it invisible. If AIs can't find it, then for all practical purposes it doesn't exist" 12.

He frames both sentences as forecasts rather than facts. They are not fully true yet, he writes, and they will only get truer. The direction of travel carries the argument: as more reading is done by systems that fetch pages, a page those systems cannot reach stops arriving in the answers people get.

He attaches no figure to the forecast. The question it leaves a writer with is which parts of their work are reachable without a subscription, and which exist mainly as an entry in a database that a summarizer has to be able to open.

Mollick: the 1979 slide saying a computer must never decide Ethan Mollick's posts on X, quoting MIT CSAIL · 1,128 likes and 60,286 views on the thread at capture · October 18–19

MIT's Computer Science and Artificial Intelligence Laboratory posted a photograph of an IBM slide from 1979. The slide reads: "A computer can never be held accountable. Therefore a computer must never make a management decision" 13.

Mollick has seen the slide before and dislikes it each time. It was already false when it was printed, he argues, because automated systems have made management decisions for decades in trading, logistics, customer service and insurance. Pretending they do not, he writes, launders the responsibility away from the people who deploy them 14.

He also turns the slide's authors against it. MIT and IBM were central to SAGE two decades earlier, which he describes as an automated system that assigned interceptors and missiles against incoming aircraft 15.

Mollick writes that he is an MIT alumnus and that his complaint is about the slide itself; CSAIL, he adds, has a large body of work on automated decision-making. His preference is that the emphasis fall on accountability and alignment work as machines take on more management decisions, instead of on a rule that forbids the decisions outright 16.

LeCun: compressed air can move a car; it cannot make a flying one Yann LeCun's posts on X · 7,530 likes and 601,812 views on the post at capture · October 18–19

At Tesla's Roadster event outside Waco, Texas, on October 15, the car jumped over a Cybercab, a Cybertruck, a Tesla Semi, a Model Y and a Model 3. Tesla's engineering vice-president Lars Moravy said ten thrusters fed by 600 liters of compressed air lifted and spun it in the air, and Tesla claimed 0–60 mph in one second 17.

Yann LeCun, a Turing Award winner who has led AI research at Meta, took the demonstration apart on October 18. Compressed-air thrusters can give acceleration in any direction, he writes, and might hold a car off the ground for a couple of seconds if you replace the passenger seat with a 400-liter high-pressure tank. Neither makes a flying car 18.

He compares it to Burt Rutan's SpaceShipOne, which made short suborbital hops and, in his words, was never going to be a path to orbit. LeCun writes that he has never criticized a real achievement, including Musk's, and owns two Teslas; his target is a marketing stunt presented as an achievement. He also states that he opposes Musk in politics 18.

The replies ran through October 19, and LeCun kept them on one question: what a demo shows, and what it is claimed to show. He called the trolling a weekend hobby and noted, in passing, that he is starting a company at 66 19.