传声·当日

Astra powered ships this week - Images 2.5 - GPT-Live-1 - Agents API - Data Agent - ChatGPT for Financial Services and yet it’s not yet DevDay. Busy plans for next week too.

TBH once you score 1600 you should unlock a second harder test that gives you a second score on top of the 1600— we should have more and higher measures of excellence But instead they ban the SAT so that everything turns into random lottery and excellence cannot be recognized https://t.co/BCdZC9tOhY

I have decided to do this after much deliberation: All local scheduled tasks live in Codex. All cloud tasks I’m porting over to Grok Bot. Makes it a clean split.

Hi Astra users. A reset and a quick update on quality issues that have been posted around. Working with some of you, we have found and fixed the following issues: - Some skills written for previous models were triggering too often or preventing the model from checking its work. - An opt-in context management experiment that could cause early stops or replies to older messages. We've disabled it. Our rough estimate is that 4-5k users were affected by this experiment. - We've also removed some badly configured engines that resulted in a measured quality degradation for a long tail of traffic flowing through them. We’ve also made some more minor improvements and things should feel significantly better across the board. More consistent follow-through, better tracking of your latest message, and better checks on the work as it’s going through the motions. The examples posted and all the users who worked directly with us were incredibly useful in helping fix things quickly. Always grateful for this incredible community. And of course, a reset is also landing by midnight today.

i emptied my house said goodbyes and moved to taipei so far so great! 💛 https://t.co/vvAbELmie8

I often think and hope that I would have had the courage that the people in United 93 had to prevent an attack on the White House. I have been thinking about 9/11 all day… I was at CMU in a lecture in Wean Hall room 7500. People were very scared because one of the planes was near Pittsburgh. So many heroes on that day.

okay this is not great https://t.co/yzMDp9dIZh

Read why here: https://t.co/qMkuBOwSKk

Excited to welcome Aidan & @ Sasha from the Git AI team to OpenAI! They are building in the open and have developed an open-source tool that helps developers understand how coding agents contribute to their codebase. Together, we’ll make it easier for businesses to see where Codex is making a difference when working through problems for individuals and teams. We’ll also keep Git AI open source and continue investing in it.

I am skeptical of "software factories." Other than for verification & testing, I don't think AI is at a point where it can self-improve a product or build a new feature end-to-end without human in the loop. Whenever I loop something overnight to build something new = if it makes one wrong assumption the whole thing becomes a waste of tokens. What products or features have been built end-to-end with a software factory and no human defining the requirements or checking the work?

“Success has many fathers, but failure is an orphan.” I never realized the gravity of this quote till I saw extremely successful VCs fight for attribution of the hot deals. And you realize this viscerally because the large exits (or markups) are genuinely rare and those are the ones you raise the next fund on (your own or part of the firm). This is why in the next couple of years you’ll see many names scrubbed (or conveniently omitted) and everyone wanting to revise history to claim some credit for the ones that eventually succeed. So many emerging GPs are facing this explicitly as the only thing they can take to their future LPs is their track record. And it’s been sad to see them not get the credit they actually deserve. As @AnjneyMidha said, eventually the new guard will shine as they will figure a way out. But in the short term, hold your founders close since they’ll be the real reference checks! (this is genuinely NOT a subtweet at any firm, just what I have been noticing across the market as I’m helping emerging managers)

Replit acquires a business entirely built on Replit. I expect first of many to come. https://t.co/hL516eAAu2

we heard feedback that it's hard to know if your skills are still working with new model releases plugin evals are here to help run `claude plugin eval init` in your plugin folder https://t.co/Q0I1ZnugDs

Tailscale’s model router is powered by Vercel AI Gateway as its underlying infrastructure. AI Gateways are the new CDNs. You could go “direct to origin”, but it’s brittle. You could DIY, but it’s painful and costly. Proud to serve Tailscale, the 🐐 networking technology company https://t.co/ERS5GKJx6D

Had to submit a patch to @trycua to make keys work reliable under Linux, dope framework overall tho!

Astra on OC in a cloud session playing Doom with CUA. Not quite AGI yet, but probably beats fly brain. https://t.co/SrJlaey3MB

You can see us on the list of tenants here: https://t.co/DuGS9PLTIw and you can see part of Kenton's story here: https://t.co/3KVbiF1Nqy

My 9/11 story (lightly updated version of a blog post I wrote in 2016 for the 15th anniversary) In September 2001, I was the young co-founder of a NYC tech startup, TripleHop Technologies.   We were building enterprise search software, using whatever AI existed at the time, building what would be today close to what a Glean does.  We had maybe 25 people on the team at the time. Pretty much all of us were at the very begining of our careers, barely out of school, unexperienced, but highly enthusiastic. Our office was located on the 53rd floor of the North Tower at the World Trade Center. We had ended up at the World Trade Center fairly randomly.  Our first office was located on Water Street, right above a 24 hour McDonald’s.  It looked more like a decrepit apartment than an actual office, occasionally didn’t smell that great, but it was cheap and good enough. But after a year or so, the team had grown, and that office had become too small, so we started looking for a bigger one. At the time, for a reason or another, there wasn’t a lot of office inventory on the market in the area where we wanted to be.   We looked for a bit, couldn’t find anything suitable, until one of us had the strange idea of inquiring at the World Trade Center, which felt like a long shot. Turned out that the World Trade Center had some smaller offices, and some availability.  In those dot-com bubble days, they also were quite eager to attract young tech startups, and were willing to offer us some preferential terms. So we went to take a look, and visited two offices in the North Tower. One was a large office at the very top of the building, perhaps the 104th floor, I don’t exactly recall. It had unbelievable views, but it really felt like taking that office would be pushing it – even with a good deal, we were just a little seed-funded startup, after all. Also, the elevator ride took forever. So instead we chose the smaller, cheaper office on the 53rd floor.   We really liked our new office, and the World Trade Center in general.  It felt like a little town, with a lot of retail and amenities below the surface, and its own subway station. We ate a bit too often at a Sbarro joint located right outside the ground floor elevator. The night before 9/11, my co-founder Renaud Laplanche and I worked very late into the evening, as we did most nights.  In true nerdy founder fashion, we played a ping pong game or two, as we did pretty much every time before calling it a night – the ping pong table being our attempt at creating a startup culture.   I left, and think Renaud may have stayed on that night to finish working on a proposal for a prospective customer. The morning after, when the plane hit the North Tower at 8:46am, I was home, about to leave for work.   Like many young tech people, we were night owls but not early birds, typically showing up at the office around 9:30am.   I had the TV on, and as about to turn it off and walk out of the door, saw the “breaking news” scroll at the bottom of the screen mentioning a plane collision, but no image.     As I had a faraway view of the World Trade Center from the window of my West Village studio apartment, I looked up and saw what seemed like a little bit of smoke at the top. I assumed the plane that had hit the tower was a small tourist plane.  We would often see them fly by the office, above the nearby Hudson River, often at altitudes that felt very close to us.  It had crossed my mind more than once that it wouldn’t take that much for one of them to hit the tower by accident. Then, like so many others, I saw those unbelievable images on TV. I called my family back in France.  Shortly thereafter, phone lines went down. The rest of the day was surreal and gut-wrenching, although I remember feeling mostly numb. I went on a mad scramble to see if everyone at TripleHop was ok. For many hours, we just didn’t know.  With no phone working, we accounted for most people via email, but a number did not reply for a while. It eventually turned out only one person from the team, Kenton Beerman, was at the office at 8:46am.   He made it out shocked but unscathed. Perhaps as a way to cope psychologically, or just because they were amazing people, I remember that the team was incredibly focused on saving the business that day.   This was obviously pre-AWS, we had our own server room in our office, but thankfully our CTO Joaquin Delgado and his team had backed everything up.  By the end of the day, we were essentially back up and running. After a few days, we found shared office space.  But business didn’t go back to usual. In addition to everything else, 9/11 was the final nail in the coffin for the economy after the dot com bubble burst.  The next couple of years were a complete nuclear winter. Thankfully, we had raised a $5M Series A (a solid amount at hte time!) a few months before 9/11, which kept us afloat. We also managed to win a few deals in Europe, which gave us a bit more cash.   Eventually business picked up strongly in late 2003 and 2004, potential acquirers came knocking, and TripleHop was acquired by Oracle in 2005. We all went on with our lives.  My co-founder Renaud reached extraordinary entrepreneurial success and defined a whole new industry. We all have families now.  I occasionally see pictures of beautiful children on my Facebook feed. I rarely talk about 9/11, and have tried to not think about it too much. It took me a very long time to go visit the 9/11 memorial downtown.   I finally made it just a few weeks ago, after a board meeting nearby, walking there almost by accident. Looking at the names engraved in the marble, I was reminded that it wouldn’t have taken much.  An hour later, a few floors lower. Or choosing that office on the 104th floor instead of the other one, a few months prior. And I feel deeply connected not just to those names, but to New York. Before 9/11, New York felt like an exciting, but ultimately foreign, city. After 9/11, I truly started feeling like a New Yorker. Everyone remembers where they were on 9/11.  Everyone has a story of some sort, many infinitely more tragic. This is just mine.

it's basically impossible to interpret evals by looking at just at the pass/fail scores these days many of the failures I see in benchmarks are due to overly strict hidden tests, in some cases the model's answer makes more sense than the expected eval result

I think the idea of a "one-person company" is overrated Yes, a person can do a lot more thanks to AI. But building something new can be a profoundly lonely experience. You need someone to brainstorm together, suffer together, celebrate together... It is extremely easy to lose motivation when you don't tie yourself to the mast together with somebody else

better benchmark scores don't tell you much about how well a model does on your real work that's why for the last 3 years @every we've done vibe checks on new models: long-form reviews based on hands-on testing on each model for our real work now we're doubling down, and getting more quantitative @hammer_mt and @nityeshaga built an internal platform for all of us to make personal benchmarks based on our real day-to-day work our vibe checks on new models used to be vibes only now we're starting to get more quantitative with checks very psyched about this :)

This one is very cool. Now you can mount Box to agent sandboxes to make it far easier for an agent to read and write files on the agent's computer. As AI agents execute critical workflows in the enterprise, they're going to need the same primitives that people have had. https://t.co/h1BW92pMtH

join the most ai pilled team on earth be at the center of the AI wave, with access to the new models, tools, and teams that matter do the best writing of your life: https://t.co/PDmrEAt4Yk

RSVP: https://t.co/WiWHBZ8pW7

How do we encourage true innovation in CA? How do we enable a two-party state? How should technology be regulated? We're hosting conversations about tech policy with decision makers at @spc this fall. @SteveHiltonx on Sept 23rd. https://t.co/sThS98c6tx

Why most enterprise AI efforts fail. 1/ Using old-school playbooks for AI The CEO picks a trusted lieutenant to lead a central AI team. That leader pulls in a group of trusted people from across the company. The problem here is the skill gap. The team structure, product patterns, how you launch and learn that got you here do not work with AI. Much of product development over the last 15 yrs has been incremental improvements. AI requires experimentation and invention. 2/ Under-investing in evals Most enterprises don't quite understand the importance of this and what good looks like here. (I have a whole evals series in the comments - check it out) 3/ Building AI for the company from the outside The central AI team declares itself a platform team and builds tools that everyone is supposed to adopt. These products are often disconnected from the actual workflows, context and judgment of the people doing the work. Begrudging adoption follows. But not meaningful productivity gains. What you should do: 1/ Hire leaders who have done successfully built AI products. You need technical/product depth plus the ability to earn trust with the people whose jobs they’re helping reinvent. 2/ Make evals a first class citizen in your AI efforts 3/ Embed your best AI builders inside the functions you want to transform. Build * with * finance , build * with * sales, build * with * support.

long-lived agents for your big ideas now in Cursor https://t.co/2sa0rX2415

Routines with budgets! https://t.co/bAweXu6MJt

Love seeing all the Astra 3D experiments on Replit! https://t.co/PC7FBalOeW

Speaker 1 | 00:00 - 00:19 The minimum fee typically on a debit or credit card transaction is about 30¢ fee flat plus the additional percentage. Arguably, it doesn't really work well for anything under a dollar. Certainly, it doesn't make sense to pay 30¢ to send a 1¢ payment. And I believe the stats we last saw on this are that about, like, 76% of the agent ecommerce transactions we're seeing are under 30¢. We don't want the AIs to be unbanked. Speaker 1 | 00:19 - 00:37 You know, if we wanna bank the AIs, they deserve financial services as well. The AI agents themselves are gonna have to have their own accounts. And so we've made some simple tools where with a single command, just prompt, you paste it into your AI agent, it can now have its own financial account. So it'll be interesting to see how the agentic economy evolves. There'll be more agents than humans in the not too distant future. Speaker 1 | 00:37 - 00:43 It kind of stands to reason that the agentic economy will be bigger than the human economy at some point in the not too distant future as well. Speaker 2 | 00:50 - 01:32 Today, know Priors. We are joined by Brian Armstrong, the cofounder and CEO of Coinbase, one of the world's most important and earliest crypto companies that has also been quite aggressive at adopting AI. Brian is also the cofounder of New Limit, which is a antiaging and longevity company doing really interesting things in biology, including certain types of high throughput screening to understand, aging, different uses of AI, etcetera. Brian is one of the most forward thinking people I know on different aspects of science, technology, art, the world, etcetera. So I was very excited to have this conversation with him, and we'll be talking about everything from what's happening in the cry