BREAKING: A mathematician at Anthropic claims to have used the company's AI to solve the Riemann hypothesis, the most notorious open problem in all of math
https://t.co/nEXAyIkvSt
I can finally share what I’ve been working on at @Microsoft. I lead Today, the product you saw all over the Ignite conference this week.
Today is a new form factor for work. It’s not a chat box it’s a Feed. While you sleep, an AI agent works in the background on your email, your calendar, Teams and your documents. In the morning, the work is already prepared: replies drafted, conflicts spotted, the messages that matter pulled to the top, the documents waiting for your review.
You don’t have to prompt it. There’s nothing to configure. You open Today, check what it did, and in a few taps you send, accept or mark as done. Most of us start the day by digging through everything that came in overnight. With Today, you start by finishing work your agent did.
We think this is what AI at work should feel like: proactive, quiet and useful, without you having to ask.
Today lives in your existing apps starting with Outlook and Copilot meeting customers in the app they use the most.
I had so much fun building it with an incredible team: @JesseProvo, @donovanso, @telmen, @justinchando, Chris Parnin, Gideon Ojo, Alex Trzyna, Christopher Miller, Diana Slaba, Kirill Varshavskiy. Thank you for the craft, the late nights, and for caring about every detail.
And none of this would have happened without the incredible support of @sumit_c, @elbio_abib, @rroslansky, @clamanna, @jacobandreou and so many others across Microsoft. Thank you for believing in it.
This is just the beginning!
P.S. i made a little film to celebrate. It starts in 1997. Watch till the end.
I think I've entered a new stage in life. In my early twenties I couldn't wait to leave home and see the world. Now I think life is about minimizing the distance between your job, your home, your gym, and your loved ones.
@guanlan Great work! Just as a reference point, I'd also love to know how the major frontier models + their own harnesses perform (e.g. claude models on claude code).
There has been a lot a lot of chatter here about getting agents to speak ASD-STE100 being nice but for observability agents BLUF and SBAR for handovers/status updates are very powerful
https://t.co/UUe56rdST5
https://t.co/ObCqqbYqes
Big congrats to the team! I remember being really impressed by the Conductor app and how quickly the team iterated.
Something weird about pure devtool startups (i.e., not SaaS businesses like Datadog/Amplitude/PostHog/Okta or hosting platforms like Vercel/Supabase/Cloudflare) is that your only realistic exit is through an acquisition.
Look at @voidzerodev, @bunjavascript, @deno_land, or even arguably @cursor_ai. Brilliant teams building brilliant products, but incredibly hard to monetize.
So the playbook is: forget about making money and go all in on building something incredible and driving adoption. Meanwhile, fundraise carefully, keep the team lean, and spend frugally so you can survive as long as possible.
You're selling your product, your team, and yourself to acquirers, not to users. Which is weird because I thought the point of a business is to make money. Instead, the winning strategy seems to be turning your startup into a resume.
This is one regret I have from my last devtool startup. If I could do it all over again, I'd focus 100% on adoption. Or better yet, I wouldn't start a devtool company at all 😂
I am very excited to share that Conductor is joining Meta!
We have some big, crazy plans for the next interface. If you liked Conductor, you’re gonna love what’s coming.
We’re super excited to work with @natfriedman and @alexandr_wang at MSL to build it. They’ve been on an absolute tear, and I’m convinced MSL has the highest talent density of all the labs.
A huge thank you to our earliest believers @aaron_epstein, @ilyasu, and @nabeel. And to everyone here who tried Conductor, sent us feedback, and prompt requests - you made this so much fun. Thank you ❤️🚂
Choo choo!
@_dylanga Curious why "compute all regexes at the module level instead of instead functions and loops"? For readability? I thought V8 already caches compiled regexes, so the performance benefit would be pretty small.
A dump of the biggest improvements to my AI coding setup lately:
1. One "god" skill
The idea is to have a single entry point for everything. I call mine /work-on. It takes a task and routes it to a "playbook", which is a list of steps for completing a task.
Now I never have to remember which skill to call. For example:
• "/work-on make a new settings page" → build-feature
• "/work-on fix the bug in issue #123" → build-bug-fix
• "/work-on cut dev build startup time by 50%" → hillclimb
See the diagram for how I set up /work-on and my playbooks. Inspired by pstack's /poteto-mode.
2. A verify skill for every repo
Each repo has a /verify-<repo> skill that runs the app locally and tests the change like a user. In my web app, it starts the app, opens a browser with agent-browser, signs in, opens the right page with the right flags, and clicks through.
I use pstack's /create-verification-skill and /maintain-verification-skill to write and maintain these.
3. Requiring screenshots on every PR
My /create-pr skill requires a before and after screenshot for every change a user can see. If they're missing, the agent takes them with the verify skill. It's a small change, but my PRs are now much easier to trust, and my coworkers understand them faster.
4. /tdd
In my experience, AI loves writing useless tests if you let it run wild. They bloat the codebase, slow down builds, and confuse future AI agents. So I adopted @mattpocockuk's /tdd skill to steer agents into writing better tests.
5. No repro, no fix
To stop AI from fixing bugs that aren't there, I added two rules to my build-bug-fix playbook:
• look for an existing fix first, and stop if one already works
• always repro the bug before fixing it
6. Babysit
Every 15 minutes, the babysit playbook checks my PR and fixes whatever blocks it: build errors, review comments, merge conflicts. It never merges. I do. I'm not at the stage of yolo-merging PRs at work (yet).
7. /correct at the end of each task
pstack's /correct skill reviews the run for mistakes: my corrections, review comments, failed builds, and hiccups a better skill would have avoided. It turns each one into a PR in my repos or in my skills so the next agent doesn't repeat it.
It's why my agent makes fewer and fewer mistakes over time.
That's it for this time!
Links:
• pstack skills (/poteto-mode, /create-verification-skill, /maintain-verification-skill, /correct): https://t.co/Ef0woSevoo
• agent-browser: https://t.co/eYHI868215
• @mattpocockuk's /tdd: https://t.co/nGmxFhEmp8
So interesting how this formula puts into numbers how I've been feeling.
Say AI speeds up coding by 10x. But if coding is just 30% of your job, you only end up with a *1.37x* speedup.
So you gotta speed up the rest: the CI, the PR reviews, the QA, the company bureaucracy, etc.
If AI makes you 10x faster at coding, why is your team not shipping 10x faster? Amdahl's law has the answer...
The law says the overall speedup is capped by the part you did not speed up. With `p` as the fraction of the work you accelerate and `s` as the speedup on that fraction, speedup is computed as attached.
Now plug in some numbers. Say writing code is 30% of the time from idea to production, and AI makes that part 10x faster. The overall speedup is 1 / (0.7 + 0.03), which is about 1.37x. Even if coding is half of the cycle, you get about 1.8x.
These numbers are illustrative, so use your own, but the shape stays the same. The rest of the cycle is still there:
- Figuring out what to build
- Aligning product, design, and other teams.
- Code reviews, which are now more than ever
- CI and CD takes the same wall time as before.
- Debugging and incidents, where code that nobody on the team wrote line by line takes longer to understand.
In fact, it can get worse than Amdahl predicts.
Faster code generation increases the load on every stage after it, so the review queue grows and the bottleneck shifts. A team that writes 10x more code with the same review capacity is just building a bigger queue, unfortunately
So if you want the team to actually ship faster, measure the whole path from idea to production and find where work waits. Then apply AI there too: review, test generation, flaky test triage, deploy safety checks. Otherwise, you have optimized the part of the system that was never the constraint.
Speeding up the typing was always the easy part :)
Incredible tour of the Xiaomi factory in Beijing. 91% of the construction of their new cars is automated. Unbelievable precision and quality. From first car in 2024 to doing 1,000 cars/day now. Thank you @shujishcn!