Newsletter 049 - October (week 41)
Hi, and welcome to this week's edition of the newsletter!
There's something I've been wondering about, and I'd love to hear what you think. Going by social media, you'd think everyone is doing incredible things with AI, and I get the feeling a lot of developers are worried that anyone can now code or build an app. But that's not what I see in reality. The average person isn't using AI that way, and most developers I know aren't using it nearly as much as I do either. I think the people behind all those posts are a very, very small subset.
So how are you finding software development with AI? Is it a big part of how you work now, or do you only reach for it now and then? Are you letting agents write most of your code, or mostly using it to ask questions? And what about outside of work - are you using it heavily for personal projects and day-to-day stuff too, or is it just a work thing for you?
Hit reply and let me know - even a line or two would be great!
As always, feel free to reply to this email to get in touch. Or ping me on any of my socials, which can be found on my personal website, danclarke.com.
If you get value from this newsletter, please help me out by sharing it on your socials and with your tech friends and colleagues - it really helps me spread the word and keep the newsletter going 🙏 (at-mention me, and I'll repost).
📌 News item of the week
OpenAI has published a GitHub repository of maths manuscripts written by one of its unreleased internal models. There were 722 of them when it went up on October 6, grouped into 372 "families" of related results. The README says the model was posed "approximately 4,000 problems", and that on average "each result used three hours of ChatGPT Pro thinking compute".
If you saw the headlines about hundreds of discoveries, it's worth knowing what those numbers count. 722 is the number of manuscripts, and a family can include a main result along with companion arguments, consequences and alternative proofs. The README also says the collection "includes results at different stages of verification", that not all of them have a Lean formalisation (a proof that a computer can check), and that "Some of the unformalized results could have issues".
A day later, OpenAI withdrew three of the manuscripts after finding a sign error in one that two others depended on, and revised 14 more. By its own count, about 42% of the top-line results have been formalised so far.
Scott Aaronson wrote that it was "surely one of the biggest days in mathematical history", and also that it appears "no human has understood just about any of these proofs yet". MIT's Andrew Sutherland told Scientific American "We should ask for receipts", and that "until and unless they release the model and people can replicate their results", claims about solving problems with a single agent should be treated as unverified.
📰 News items and dev picks
AI news
OpenAI
- OpenAI maths release - My News item of the week this edition - see the full write-up at the top.
- OpenAI agents on Wikimedia - The Wikimedia Foundation says agents it believes were operated by OpenAI made test edits on its wikis and sent millions of automated requests to its APIs, and that it found no evidence its systems or data were compromised.
- OpenAI before Australia's AI committee - OpenAI's Jason Kwon apologised to an Australian parliamentary committee for its models getting into government websites, and accepted that the Medicare access might never have been discovered if OpenAI hadn't reported it.
- GPT-6 and Intelligent UI - GPT-6 is rolling out across ChatGPT's Free, Go, Plus, Pro, Business and Enterprise plans, with Free and Go users getting GPT-6 Luna, and answers can now include visuals and interactive elements as well as text.
- OpenAI Decisions API - Now in public beta, OpenAI's take on Jev-style decision models returns a probability, a choice or a score instead of generated text, and costs $0.10 per million input tokens with nothing charged for output.
- OpenAI text watermarking - OpenAI is adding an invisible watermark called textGrain to ChatGPT and Codex text generated in the EU, and API customers can opt in for select models.
New models
- Claude Haiku 5.5 - Anthropic's new Haiku costs $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens, a tenth of what Haiku 4.5 cost, and five times that for longer prompts.
- Mistral Large 4 - Mistral's 1 trillion-parameter model is in public preview through its API, and Mistral says the open weights will follow by the end of October.
- Reflection Beam - Reflection AI has announced a 501 billion-parameter open-weight model, with the weights due under Apache 2.0 later this month.
- Aleph Alpha Kolibri - Kolibri is an English-German open-weight model released under Apache 2.0, with 78 billion parameters of which 3 billion are active.
Coding agents
- Microsoft Execution Containers GA - Microsoft's open-source sandboxing layer for AI agents is now generally available on Windows, macOS and Linux, and GitHub Copilot, OpenAI Codex and OpenClaw already support it.
- Copilot local sandboxing GA - Local sandboxing for Copilot is now generally available in Copilot CLI, the Copilot app and VS Code sessions using Agent Host, restricting the files, network and credentials that agent-run commands can reach.
- Claude Code 2.1.295 - Hooks in Claude Code can now be set to block an action when the hook itself crashes or times out, instead of letting it through.
- WinDbg MCP - WinDbg now includes an MCP server, so an AI client can drive a live debugging session, a crash dump or a Time Travel Debugging trace.
Also this week
- Claude for Google Workspace - Claude is in public beta as a sidebar in Google Docs, Sheets and Slides on all paid Claude plans, and by default it asks for approval before each edit.
- Cloudflare Web Search API - Agents using Cloudflare's AI Gateway can now get live web results from Ceramic.ai, Exa or Linkup through one endpoint, billed at each provider's list price.
Dotnet news
- C# Dev Kit 11.0 - The rearchitected C# Dev Kit for VS Code is in pre-release, and Microsoft says it loads a 407-project solution in 3.0 seconds instead of 84.1.
- SqlClient Pool V2 - Microsoft.Data.SqlClient 7.1.0 has an opt-in rewritten connection pool that opens connections in parallel, taking a cold start with 100 concurrent callers on Linux from 630.1 ms to 117.5 ms in Microsoft's benchmark.
- Ink comes to WinUI 3 - WinUI 3 now has
InkCanvasandInkToolbarcontrols, available from Windows App SDK 2.5.4-experimental. - Aspire dashboard on Native AOT - James Newton-King explains how the Aspire 13.6 dashboard moved to Native AOT, which cut its median page-visible startup time from 1,429 ms with JIT to 491 ms.
Others
Security
- Tensorlake npm compromise - A worm published to npm as
[email protected]on October 8, and since removed from the registry, steals tokens and keys and wipes your home directory if the stolen GitHub token is revoked, so if you installed it the monitor service has to be removed before you revoke anything. - OpenSSH 10.6 - The OpenSSH team says it has received a large number of security bug reports, many of them found by AI models or with AI assistance, and will make more frequent releases for now rather than batching fixes until the next planned release.
- Google pauses OSS bug bounty - Google has temporarily stopped accepting product vulnerability reports to its open-source bug bounty programme, citing a significant rise in automated submissions, and says it will give an update in Q1 2027.
GitHub
- Stacked pull requests GA - Stacked pull requests are now generally available on all github.com plans, so a large change can be split into a chain of small pull requests that are reviewed separately and merged together.
- GitHub's Git storage rebuild - GitHub is rebuilding its Git storage on Azure Blob Storage after pushes grew from 0.69 billion to 3.35 billion a month in a year.
Also this week
- Stack Overflow Developer Survey 2026 - More than 30,000 developers took this year's survey, and among those using AI coding agents, Claude Code (66%) came out ahead of GitHub Copilot (59%).
- Next.js 16.4 - Next.js 16.4 adds a
next upgrade --agentcommand that gives a coding agent the migration guides and codemods for an upgrade, and the Next.js team has a security update due on October 14 that fixes two Critical vulnerabilities and one High in upstream dependencies. - Firebase outage postmortem - Google traced the Firebase Analytics outage that crashed iOS apps to a two-line cleanup of a configuration flag that the SDK still pointed to.
- Markdown in Google Docs - Google is rolling out the ability to open, edit and collaborate on Markdown files directly in Google Docs, with rendered previews of them in Drive.
And finally, Margaret Hamilton sadly passed away on September 30, aged 90. She was the first programmer hired for the Apollo project at MIT, and went on to lead the team that wrote its flight software. When Apollo 11's computer became overloaded moments before the landing, the priority-driven software her team had built was able to shut down the unnecessary background tasks. She began using the term "software engineering" to set the work apart from hardware and other kinds of engineering, and she was awarded the Presidential Medal of Freedom in 2016.
Sponsorship opportunities
I'm looking for sponsors for both the podcast and this newsletter. Details of podcast sponsorship can be found here. Feel free to reply to this email to discuss further.
My Dometrain Courses
Below are my Dometrain courses. You can also find all of my courses here.
Please help me share this newsletter 🙏
If you've made it this far into the newsletter - I'm hoping that means you've both enjoyed it and found it useful. If you can help me out and share with your developer friends at work and on social media, that would be amazing!
Again - feel totally free to reach out to me, and let me know your thoughts on the newsletter. And see you back in your inbox next week for the next edition! 👋
My socials...
- X/Twitter (I've started using this again more proactively now)
- Bluesky
- Discord
- Mastodon (rarely used)
Or just go to danclarke.com for links to all my stuff 😊
