← 2026-05-04

Daily Edition

2026-05-05

2026-05-06 →

AI Builders 日报 — 5月5日

追踪 AI 领域真正在做事的人,而不是空谈者。

今日思考

Sam Altman 今天发了至少 6 条关于 GPT-5.5 的推文,从 rate limits 争议到模型能力边界,无一不在回应社区疑虑。值得注意的是,他主动征集"高 token 消耗"的用例——这说明 OpenAI 在为下一代长上下文产品铺路。与此同时,Replit Agent 单日 50 万项目的爆发式增长、Garry Tan 在斯坦福分享"6 人团队现在能做到 10M 收入",都在印证同一件事:AI 能力的上限已经不是问题,问题是组织能不能跟上这个速度。


产品与发布

Replit Agent

周末免费试用引发大规模涌入,Replit 单日产出 50 万个项目,有用户消耗了价值 1 万美元的 Agent 工作量,有人一天探索了数百个创业方向。基础设施峰值承载了平时 4 倍的负载。faviconx.com

GBrain v0.27

Garry Tan 宣布 GBrain 新版本正式支持非 Anthropic 和非 OpenAI 的 embedding 提供商,以及更多 LLM 提供商。多模态 embedding 和深度图片 OCR / EXIF 提取功能即将上线。faviconx.com

Pomelli Catalog(Google Labs)

Google Labs 推出 Pomelli Catalog,允许用户将产品或服务信息同步到 Pomelli,平台会基于品牌信息生成个性化营销素材和高质量商品图。免费使用。faviconx.com

GPT-5.5 Instant

OpenAI 向所有 ChatGPT 用户推送 GPT-5.5 Instant 更新,重点提升智能水平、图片理解和事实准确性,同时将写作风格调得更直接、更简洁。Sam Altman 本人表示这是"相当大的升级",特别提到速度、智能、个性和记忆/个性化功能组合在一起,带来了"超越各自之和"的体验。faviconx.com


观点与判断

Sam Altman(OpenAI 创始人)

  • Codex 的 rate limits 是故意设计的"蜜月"? 有人在推上质疑 Codex 的限制过低,怀疑是先用低价吸引用户迁移后再提价。Altman 直接回应:"我们的模型效率很高,特别是考虑到其能力水平。祝 codexing 愉快。"并补充:"冲 rate limits 来,留下来的原因是最好的模型。"faviconx.com

  • GPT-5.5 在 Codex 里做非编码任务也很强 Altman 表示自己在 Codex 里使用 5.5 时不断惊讶于它的能力边界——"我经常假设它做不到某件事,但很多时候结果是令人愉悦的意外。"faviconx.com

  • 寻找 GPT-5.5 的极限用例 Altman 在公开征集"用 5.5 做出之前模型做不了的事"的案例,尤其是高 token 消耗的场景。faviconx.com

Garry Tan(Y Combinator 总裁 / GBrain)

  • OpenClaw 遭遇僵尸 Worker 问题 Garry 在运行 OpenClaw 时遇到严重的僵尸进程问题,询问 tini 是否是正确解法。这是生产级别的工程问题。faviconx.com

  • Y Combinator × GBrain Hackathon 定档 5 月 16 日 Garry 宣布将在 YC 举办 GStack × GBrain 黑客松,12 小时构建并发布真实项目,奖品包括 YC 面试资格和与 Garry 一对一的工作交流。报名地址:faviconevents.ycombinator.com faviconx.com

Peter Yang

  • 三阶段论:编码→知识工作→个人 Agent Peter 提出 AI 落地的三段论:"编码是第一个前沿,知识工作是第二个,个人 Agent 是第三个。"faviconx.com

  • Mercury VP 用 AI 当教练 Ryan(Mercury VP)构建了一套基于 Claude Code 的系统,将会议记录与绩效评估交叉比对,AI 直接告诉他"你在这个会议里做的正是你绩效报告中提到需要改进的地方"。Peter 评价:"AI 当教练的效果出人意料地好。"视频链接:faviconyoutube.com

X / Twitter

45
petergyang
petergyang @petergyang
I want to get my 8 year old to start building stuff with agents that she actually can share with her class and teachers. Any ideas?

Maybe she can make her first dollar online too.
garrytan
garrytan @garrytan
Retweeted
在悉尼和稀泥 在悉尼和稀泥
Re 5/ YC CEO Garry Tan:为什么开源GBrain、跑自己的栈?
.@garrytan 的逻辑很硬:intelligence explosion意味着掌控自己的context比以前任何时候都重要。Own your prompts, own your data = 真正的思考自由。
⚠️ 反直觉但值得想:你只用ChatGPT/Claude默认体验时,你的思维方式也在被默认设置悄悄塑造。
https://x.com/garrytan/status/2051110206466302136
sama
sama @sama
we love you too!

Arav Jain: WHAT THE HELL I LOVE YOU @OpenAIDevs
10X RATE LIMITS ON CODEX IS FRICKING CRAZYYYYYYYY

sama
sama @sama
pretty excited for voice models to get great

its interesting to watch how people are already starting to change the way they interface with AI
garrytan
garrytan @garrytan
Retweeted
Gabriel Jarrosson Gabriel Jarrosson
Founders complain about YC taking 7%.
It's the wrong math.
You're not giving away 7%.
You're buying into a system that has produced Airbnb, Stripe, and Coinbase.
The outcomes from YC companies run at least 4x higher than the market average.
7% of something enormous beats 100% of something average.
The most expensive decision isn't joining YC.
It's not joining because of the cap table math.
petergyang
petergyang @petergyang
Met the demo god himself today @romainhuet
petergyang
petergyang @petergyang
Coding is the first frontier.
Knowledge work is the second one.
Personal agents are the third.
amasad
amasad @amasad
Replit helped an entrepreneur find investors and land meetings!

Said Sadik: This true. I built the pitchdeck in @Replit , it was awesome, but it didn't stop there! I told it I need to find investors to send it to, it gathered around 1897 investors, and walked me the process of buying a postmark plan and apollo plan. I bought them both, enriched data in
garrytan
garrytan @garrytan
GBrain v0.27 just dropped - as promised, you asked for support for lots of other non-Anthropic and non-OpenAI embeddings and LLM support, so here it is.

Multi-modal embeddings and deep photo OCR, description, and EXIF extraction coming shortly after that.
garrytan
garrytan @garrytan
Retweeted
Kevin Bass Kevin Bass
I have conducted the most comprehensive public records audit of any Congressman in the history of the United States.
That audit was conducted on Congressman @RoKhanna.
This audit has exposed shocking ethical lapses and potentially criminal behavior by Congressman Khanna.
I am filing a 239-page ethics complaint, including 30 evidentiary exhibits, with the Office of Congressional Conduct (OCC), to be followed by complaints to the House Ethics Committee and the Department of Justice (DOJ) in the coming days.
Besides being based on an extremely comprehensive public records audit, the complaint is the first of its kind in another way: the factual basis of every single specific claim in the complaint is fully verifiable and reproducible by anyone with a computer.
Attached to this post is a link to the GitHub Release containing the complete reproducibility kit. Anyone with Python 3 and the GitHub CLI installed can download it and run a single command — `python http://welcome.py` — which walks them through the analysis at whatever verification depth they pick:
1. A 30-second offline check that every body figure derives from the bundled snapshots;
2. A primary-source spot-check that re-fetches the underlying records from the House Clerk and IRS and confirms the bytes match;
3. An OpenTimestamps proof that the package existed at publication time and wasn't backfilled; and
4. An opt-in path that lets the reviewer re-run the OCR pipeline themselves against the primary-source PDFs.
This means that any person in the world can confirm for themselves that all statements made in this complaint are fully reproducible and true.
---
The complaint asserts the following:
Representative Ro Khanna is a Democratic congressman from California's 17th District (basically Silicon Valley). He has been in Congress since January 2017. He is currently in his fifth term.
Khanna has done six different things wrong.
Each one is bad enough to investigate on its own.
Together, they are very bad.
His family's stock trades line up suspiciously with the committees he sits on, the donors who fund him, and the votes he takes.
That's bad.
Khanna's household made between $15 million and $108 million from these trades, with a middle estimate of about $61 million.
The estimate cannot be made any better than this. The disclosure forms provide only disclosure "bands". Precise amounts can only be determined with subpoena power.
But we do have one hard number:
Compared to just buying a basic stock-market index fund, his family beat the market by about $28 million.
$28 million.
The complaint says that Congressman Khanna should pay this money back.
Now, how the trading actually works in this household is important because it helps us to understanding everything else, so I will explain that now.
Khanna himself has filed 114 reports with the House Clerk listing every trade his household has made. Those reports cover 37,238 individual trades. That's a huge amount. Most members of Congress don't trade nearly that much.
But here's the kicker.
Almost none of those trades are in Khanna's own name.
99.997% of them are listed as belonging to either his wife (Ritu Ahuja Khanna) or his dependent child.
That's basically all Khanna trades. A massive volume.
Yet virtually none in his own name.
Curious.
Khanna has publicly said this is fine because the trading is done through what's called a "separately managed account" or "blind trust", meaning a broker or trustee makes the decisions without telling him.
If that were true, he'd be off the hook because he wouldn't know what was being bought or sold.
The complaint says that's not true. When you read his official financial disclosure form (the one he signs every year), it shows:
> No separately managed account
> No blind trust
> No third-party broker handling the actively-traded stocks
Instead, the trades come from about a dozen family trusts (the Ritu Ahuja 1994 Trust, the Ritu Ahuja 1995 Trust, the Ahuja Children's Trust, etc.).
These are family-controlled entities.
Whoever's making the trade decisions is a family member. His wife or his child. (Put another way: his "wife" or his "child".) Not an outside professional.
Uh oh.
The "I didn't know what my spouse was trading" defense doesn't work. Nothing on the official paperwork supports it.
Think about it.
Do you think Khanna and his wife sit around and his wife is just buying Palantir stocks, while, by coincidence, Khanna sits on the defense tech committee?
And they don't talk?
That's the framework. But it gets a whole lot worse.
Because the complaint isn't undergirded merely by this speculation. But by hard evidence.
The complaint makes six specific allegations, or "counts".
---
COUNT 1: Filing trade reports late
This sounds like a technical detail, but it is not. It is the pattern of misbehavior that enabled everything else.
When a member of Congress, their spouse, or their kid makes a stock trade worth more than $1,000, they have to report it within 45 days. That's the STOCK Act, passed in 2012. Each late report costs at least $200 in fines.
Out of about 36,000 auditable trades made by Khanna, 624 were filed late.
The worst one was 358 days late -- almost a full year. A trade in HUMANA stock made in October 2023 wasn't reported until November 2024.
The complaint provides a calculation of how Khanna fares compared to other Congressmen in terms of how often he is late in filing.
Khanna's rate of late filing (1.74%) is better than most members of Congress. The average House member is late on 10% of trades.
So if you measured just the percentage, he'd look fine.
But here's where things get crazy.
The complaint uses a special "composite score" that combines (1) how much money is involved, (2) how late, and (3) how many trades.
By that score, Khanna ranks in the top 7% of the entire House.
This means that Khanna's late filings expose more dollars to delayed disclosure than 93% of members.
A late report means the public can't see what a member of Congress is buying or selling at the time it happens.
By the time it's disclosed, the value of the inside information is gone.
The late filings are not hitting Khanna on a technicality.
They imply that the entire system designed to prevent insider trading in Congress is broken inside Khanna's office.
The 45-day disclosure rule is not a paperwork deadline. It is the security camera. It is the only mechanism that lets the public see what a Congressman is buying while the trade still matters -- while the bill is still being debated, while the FDA decision is still pending, while the news is still fresh.
When Khanna files 358 days late, the camera is off. By the time anyone sees the trade, the moment has passed. The witnesses have moved on. The dots cannot be connected.
A few late filings is a paperwork mistake. 624 of them, on a household making 37,000 trades, in the exact industries Khanna's committees regulate, is a system.
It is Khanna's system. It is how he does his dirty work.
And it is the system that lets every other count in this complaint happen in the dark.
Until now.
The complaint asks for:
1. Civil penalties for the late filings.
2. A requirement that Khanna set up an actual qualified blind trust going forward.
3. An Ethics Committee finding under House Rule XXIII that the absolute-count and composite-score chamber rankings reflect conduct that does not reflect creditably on the House.
---
COUNT 2: Buying defense stocks right before defense bills pass
Members of Congress can't trade based on inside information they got from doing their congressional job (the STOCK Act, sections 3 and 4).
Khanna sits on the House Armed Services Committee, which writes the giant yearly defense bill (the NDAA).
And across four different years, his household bought stock in big defense contractors (Boeing, Lockheed Martin, Northrop Grumman, Raytheon, etc.) right before the NDAA passed:
> 7 defense stock buys 12 days before the 2018 NDAA
> 4 defense stock buys 4 days before the 2021 NDAA override
> 1 Palantir buy 13 days before the 2022 NDAA
> 2 Raytheon buys 2 days before the 2024 NDAA
Khanna publicly voted NO on 12 of 13 of these NDAA passage votes.
So he's saying "I oppose this bill" with his vote.
But his family is buying stock in the companies that would benefit from it passing.
That, of course, is insane.
The complaint argues this is the worst version of the conflict:
Khanna gets the political credit for opposing the bill.
Meanwhile, he makes money from insider knowledge from sitting on the Committee, knowing it would pass anyway.
In addition.
Khanna sits on a committee that oversees defense contracts. The data analytics company Palantir got $4.88 billion in federal contracts during his time in Congress.
On at least nine separate days, Palantir got a federal contract AND Khanna's household bought Palantir stock the same day.
One of these was a $19 million Air Force contract on May 10, 2022: the same day his dependent child's account made six separate Palantir trades.
Khanna's defense trades made about $5.4 million in profits beyond what the broader market did, suggesting that Khanna was using his insider knowledge -- through the intermediary of his dependent child -- to beat the market.
What the complaint asks for:
1. Send to House Ethics.
2. Send to DOJ for possible criminal charges.
3. Force Khanna to give back the $5.4 million.
---
COUNT 3: Buying drug company stocks right before government drug actions
COUNT 3 is the same as COUNT 2, except healthcare stocks instead of defense stocks.
Yes, Khanna is doing the same thing across stock classes. Of course.
Khanna sits on a committee that oversees the agencies regulating drug companies (HHS, CMS, FDA). The complaint identifies 14 different government drug-pricing actions between 2017 and 2024 where Khanna's household made pharmaceutical-company trades within 14 days of the action.
1,244 pharmaceutical-sector trades clustered within ±14 days of these events. That's chamber rank 1 of 66 House members, 14 times the chamber 95th-percentile.
The biggest example: On August 2, 2024, Khanna's family made 286 trades in a single-day rebalance.
Hidden inside was simultaneous trading in four of the nine drug companies (AbbVie, Amgen, Johnson & Johnson, Merck) whose drugs were going to be on the government's negotiated-price list.
That list was published 13 days later, on August 15, 2024.
It was confidential and not yet public on the day of the trades.
But Khanna had insider access to the list. And made the flurry of trades that aligned with it at precisely the right time.
Two other "conflict triangles" the complaint highlights:
1. Palantir (already mentioned in Count 2): Khanna chairs the China select committee and is a top member on the cyber subcommittee. Palantir is a defense tech company affected by both. His family has done 29 Palantir trades and gotten $22,700 in donations from Palantir's chief operating officer.
2. Nvidia: In 2024, Khanna's family donated 10,076 shares of Nvidia stock (worth about $1.67 million when given, much more later as the stock soared) to a family foundation. In the same year, he voted NO on a chips bill, voted YES on four China-policy bills, and continued chairing the China committee. This is the committee that has the most influence over Nvidia's massive AI chip business.
3. The Goldman Sachs margin loan setup: Across 2017-2019, Khanna's spouse had two simultaneous Goldman Sachs margin loans (basically borrowing money against stocks to buy more stocks).
Each loan was labeled as belonging to a family trust ("Ritu Ahuja 1994 Trust" and "Ritu Ahuja 1995 Trust"). This same Goldman Sachs is also the broker for a sophisticated short-volatility options trading program in the spouse's account, and Goldman employees have donated about $48,000 to Khanna over the years.
You can't run an options trading program on a margin account passively; somebody (the spouse) has to authorize each trade.
What COUNT 3 asks for: Same as COUNT 2:
1. Send to Ethics.
2. Send to DOJ.
3. Force Khanna to step away from CMS, FDA, and defense matters pending investigation.
---
COUNT 4: Khanna's family trades line up with insider events at the issuer level — same-day SEC filings and same-day insider trades
The single sharpest count in the complaint.
The legal hook is the STOCK Act §§ 3-4, codified at 15 U.S.C. § 78u-1(g) — the federal statute that extends Rule 10b-5 insider-trading prohibitions directly to Members of Congress who trade on material non-public information acquired through their legislative or oversight duties.
Khanna's household trades are not just suspicious because of how many they are. They are suspicious because they happen at very specific moments.
Two examples:
> 186 of his household's trades happened on the same calendar day that the company in question filed important news with the SEC (Form 8-K — the disclosure form companies file for material acquisitions, executive changes, regulatory actions, and the other news events the SEC requires public companies to disclose immediately).
> 86 of his household's trades happened on the same calendar day that a named officer at the same company (CEO, CFO, board member) was buying or selling their own stock in the same direction.
On each of these patterns, Khanna ranks at the top of the entire House:
> Same-day-8-K count: rank 1 of 96 House Members. 4.3 times more than the second-place Member.
> Same-day-aligned-insider count: rank 3 of 156 House Members.
The complaint does NOT allege that Khanna's RATE of same-day-8-K trading is exceptionally high. As a percentage of his trades, his same-day-8-K rate is 5.4% — which is above the chamber median (4.5%) but inside the normal band. The complaint discloses this candidly, up front, to pre-empt the inevitable "his rate is in-band" defense.
The argument is about ABSOLUTE count combined with ticker-specificity: the same-day intersections concentrate on companies in sectors his committees regulate.
These two findings join two more from Count 3:
> 4,595 pharmaceutical trades within 14 days of FDA Advisory Committee meetings. Rank 1 of 66 House Members. 6.1 times the second-place Member.
> 1,244 pharmaceutical trades within 14 days of CMS rulemaking events. Rank 1 of 66 House Members. 14 times chamber P95.
Across four independent issuer-event and regulator-event substrates — SEC 8-K filings, named-officer Form 3/4/5 filings, FDA Advisory Committee calendar, CMS rulemaking calendar — Khanna's household ranks first or third by absolute count. The four substrates are independent: different agencies, different filer classes, different denominators. The convergence is structurally inconsistent with portfolio management that doesn't draw on contemporaneous information advantage.
The complaint asks for:
1. Ethics Committee referral for full investigation.
2. DOJ referral for criminal review under 15 U.S.C. § 78ff (Exchange Act criminal penalty) if any single windowed trade reflects willful use of material non-public information.
3. Disgorgement under STOCK Act § 9 of any profit attributable to same-day-issuer-event or same-day-officer-aligned trading.
4. A House Rule XXIII finding that the four-substrate convergence reflects conduct that does not reflect creditably on the House.
---
COUNT 5: Ex-government officials who became lobbyists are donating to him
The law says that federal officials who leave government can't immediately go lobby their old agencies. Various waiting periods apply, and the lifetime ban (18 U.S.C. § 207(a)(1)) prevents them from ever working on the same specific matters they personally worked on in government.
Yet, five former federal officials, who all later became registered lobbyists, donated to Khanna's campaign. Each one's old job lines up with what they're now lobbying about:
1. Chris Israel. Former Deputy Assistant Commerce Secretary. Now lobbies for tech and pharma companies (Qualcomm, AbbVie, PhRMA). Donated $1,000 (one $500 check was refunded within 24 hours).
2. Arshi Siddiqui. Former senior staffer to Speaker Pelosi. Now a partner at Akin Gump, lobbying on Armed Services issues for RTX (Raytheon) and Honeywell. Donated $2,000.
3. Francisco Sanchez. Former Obama Commerce Department Under Secretary for International Trade. Now lobbies on international trade issues. Donated $1,250.
4. Kevin Batteh. Former CFTC counsel. Now lobbies on CFTC and DoD issues for Citadel and D.E. Shaw. Donated $1,000.
5. Robert Taylor. The most damning case. Former Deputy Assistant Secretary of Defense for Senate Affairs. Now lobbies for Boeing, BAE Systems, Aerojet Rocketdyne, Textron — the exact defense contractors his old job covered. Donated $1,000 (NOT refunded). Khanna sits on Armed Services.
Their employees too: The companies these lobbyists work for collectively gave $365,140 across 264 individual contributions to Khanna.
Khanna says he doesn't take corporate PAC money. But the corporations' executives give to him personally.
Lobbyists are required to disclose their political contributions. Two of the five lobbyists hid the Khanna donations from their required reports. Robert Taylor's case is the worst: he affirmatively certified "I made no contributions" while a Khanna donation was sitting in the period.
The complaint asks for:
1. DOJ referral for the lifetime-ban review (especially Robert Taylor).
2. DOJ referral for Taylor's allegedly false lobbying disclosure.
3. FEC audit.
COUNT 6: The Ahuja family foundation and a missing rental property
Three problems.
PROBLEM 1: Khanna's family foundation isn't disclosed as a spouse asset
Remember how 99.997% of the trades made by Khanna are made either through his spouse or his child?
His wife's Ahuja Charitable Foundation is a $45 million private family foundation. His wife Ritu Ahuja Khanna, is:
> A named trustee every year from 2018 through 2024 (according to the foundation's own IRS filings)
> A substantial contributor for tax years 2022, 2023, and 2024 (also per IRS filings)
The foundation owns massive amounts of stock in defense companies (Honeywell, L3Harris, TransDigm, Boeing, GE Vernova) and healthcare companies, again exactly the sectors Khanna's committees oversee.
Khanna's annual financial disclosures don't mention the foundation as a spouse-held asset at all. And they don't mention his wife's trustee role. Federal ethics law (5 U.S.C. § 13104(d)(1)(A)) requires members to disclose their spouse's income from nonprofit positions where the spouse has decision-making power. The complaint says the Ethics Committee should decide whether this should have been disclosed.
Now, in 2024, Khanna's wife "donated" 2,821 shares of Nvidia to the Foundation, and the related Ahuja family trust donated 7,255 more shares
This was a combined 10,076 shares of Nvidia worth $1.67 million at donation time (much more later).
This happened the same year Khanna voted on multiple chip and China bills and continued chairing the China committee.
PROBLEM 2: A rental property in Dover, Delaware is missing
In tax year 2021, Khanna disclosed a $100,000-$250,000 mortgage from "First Bank of Wilmington, Delaware" tied to a Dover, Delaware rental property.
But across ten years of disclosures (2014-2023), the Dover, Delaware property itself never appears as an asset. Federal law says any rental property worth more than $1,000 has to be disclosed.
And here's the killer: Every other rental property the household owns (Cincinnati OH, Denham LA, Walton Hills OH, Harahan LA, an NY condo, Walton OH) is correctly disclosed both as an asset AND with the rental income.
Only Dover, Delaware is missing on both sides. So the household clearly knows how to fill out the form. They just didn't for this one property.
Why?
What's special about that property?
The public deserves to know if Khanna is hiding something.
PROBLEM 3: Margin loans and options trading prove there's no blind trust
Across 2017-2020, Khanna's spouse had Goldman Sachs margin loans (borrowing against stocks). At the same time, the household was running a sophisticated options trading program. They were writing PUT options on the spouse-owned account.
Under brokerage rules, writing options on a margin account requires personal customer authorization. You can't run an options program with a passive blind trust.
The "I have no idea what my spouse is trading" defense is impossible.
Khanna knew. And he was breaking the rules.
The complaint asks for:
1. Ethics Committee review of the foundation question.
2. Per-year corrective filings on the Dover property.
3. Civil penalties.
4. A possible "honest services" fraud referral if the Ethics Committee finds intentional concealment.
---
How much money Khanna made
> $61 million in profits the family made from these trades (middle estimate)
> $28 million of that is "alpha" — money beyond what just buying an index fund would have earned
> 41% of those profits ($25.2 million) came from trades made within two weeks of an event Khanna could have known about because of his job
> The complaint asks for that money to be paid back (called "disgorgement") under STOCK Act penalty rules
What the complaint asks
1. The Office of Congressional Conduct should investigate and refer the case to the House Ethics Committee for a real investigation
2. Parts of it should go to the FEC for the LD-203 lobbyist-contribution-disclosure compliance audit
3. Parts of it should go to the DOJ for possible criminal review (insider trading under 15 U.S.C. § 78u-1(g) and § 78ff; lifetime lobbying ban violations under 18 U.S.C. § 207; false statements on lobbyist disclosure filings under 18 U.S.C. § 1001 and 2 U.S.C. § 1606)
4. Khanna should set up an actual blind trust to prevent this in the future
5. He should recuse himself from CMS, FDA, and defense matters while it's being investigated
6. The roughly $28 million in market-beating profits should be returned
garrytan
garrytan @garrytan
Retweeted
Jeffrey Emanuel Jeffrey Emanuel
See the announcement today from @brian_armstrong at Coinbase in the attached pic.
Managing fleets of agents effectively is the most high-impact thing you can do now for your company and for yourself personally.
I have the most advanced set of open-source, provider-agnostic agent tooling to do all that at scale:
MCP Agent Mail for inter-agent messaging and file reservations (avoid the git worktree anti-pattern and merge hell).
beads_rust for agent-intuitive task management that works well across all harnesses.
bv for triaging tasks based on the graph structure of task dependencies (maximize development velocity through de-bottlenecking).
ntm for doing agent orchestration (creating and managing swarms of Claude Code and Codex instances).
dcg for preventing the agents from doing destructive things.
cass for letting agents instantly search all session history globally for any agent harness
ubs for polyglot “super linting” and bug finding.
You can find all of these on my GitHub:
https://github.com/Dicklesworthstone
Or install all of them automatically on a Linux cloud server using the acfs setup wizard, available for free at
https://agent-flywheel.com/
And see my complete guide explaining how to do planning with these tools, which is the key to achieving high code quality:
https://agent-flywheel.com/complete-guide
And then to make these tools as effective as possible, I have a complete suite of powerful and sophisticated agent skills that are deeply integrated with the tooling available at
https://jeffreys-skills.md/
The skills are $20/month. You don’t need to use them but they’re absolutely worth it if you want to get the most out of the tools and the agents in general.
Why do this now?
What’s happening at Coinbase is going to happen soon at every well-managed technology company within the next year.
The writing is on the wall. This isn’t alarmist fear mongering. This is the inexorable reasoning of the market and capitalism grappling with these irresistible economic forces.
Don’t wait and become a statistic, facing a structurally difficult job market without the skills that all well-positioned employers are going to be looking for.
Take action NOW to master these technologies and empower yourself.
garrytan
garrytan @garrytan
Retweeted
Alfred Lin Alfred Lin
http://x.com/i/article/2051350347730599936
ylecun
ylecun @ylecun
Retweeted
Micah G. Allen Micah G. Allen
The brain does not predict every upcoming word during language comprehension with maximum precision, contradicting the assumption that next-word prediction serves as the central computational objective of the human language system.
https://www.nature.com/articles/s41593-026-02272-6
garrytan
garrytan @garrytan
Retweeted
Patrick OShaughnessy Patrick OShaughnessy
My guest today is Brian Chesky (@bchesky), founder and CEO of Airbnb and one of the great consumer founders of the last 20 years.
Paul Graham coined "founder mode" based on Brian's experience running Airbnb. This conversation is about what comes after it, what he calls AI founder mode, and how it will force founders to focus even more on the details.
We talk about his eleven-star exercise for finding product market fit, why your first hire should be a recruiter, and why Airbnb's $100B IPO became one of the saddest days of his life.
Brian still comes across like the 17 year-old at the Rhode Island School of Design (RISD) who picked to study industrial design. His heroes are all artists. Da Vinci, Van Gogh, Walt Disney, and Steve Jobs, all of whom were working the week they died because they loved what they did.
Rick Rubin taught him that an artist is only an artist when they make things for themselves. Now Brian believes AI is the opportunity for all of us to do the same.
Enjoy!
Timestamps:
1:00 Studying Industrial Design
11:33 AI Founder Mode
17:02 Lack of Consumer AI Companies
22:10 Small Teams and Focused Problems
30:52 The Evolution from Founder to CEO
38:13 The 11-Star Experience
41:07 AI as a Canvas for Creativity
48:17 Detaching from Success
53:12 Founder-Led Moats
58:34 The Next Chapter of Airbnb
1:03:08 What Endures in the Age of AI
1:06:43 Lessons from Bodybuilding
1:10:20 The CEO's No. 1 Job
1:17:01 Activating Talent
1:20:39 The Kindest Thing
sama
sama @sama
we have very efficient models, especially for their capability level

happy codexing

kitze: yo, i'm actually worried. codex limits are genuinely insane so it's sus af ..

i feel this is an intentional move for a honeymoon period until we get over the claude → codex migration and then we get rugpulled hard
sama
sama @sama
come for the rate limits, stay for the best model
garrytan
garrytan @garrytan
OpenClaw is giving me the craziest zombie worker problem

Has anyone solved this properly?

Is tini the correct solution?
petergyang
petergyang @petergyang
"It's really been surprising how good of a coach [AI] is."

@rywiggs (Mercury VP) built a Claude Code system that reviews his meeting transcripts from @meetgranola and cross-references them against his performance review and coaching feedback.

"It tells me, hey, in this meeting, you were doing this exact thing from your performance review. It keeps me honest at a much faster frequency than any other system can."

📌 Watch Ryan talk more about how to set up an AI second brain here: https://www.youtube.com/watch?v=KzqpK1uCczw&t=790s
petergyang
petergyang @petergyang
Retweeted
Peter Yang Peter Yang
"It's really been surprising how good of a coach [AI] is."
@rywiggs (Mercury VP) built a Claude Code system that reviews his meeting transcripts from @meetgranola and cross-references them against his performance review and coaching feedback.
"It tells me, hey, in this meeting, you were doing this exact thing from your performance review. It keeps me honest at a much faster frequency than any other system can."
📌 Watch Ryan talk more about how to set up an AI second brain here: https://www.youtube.com/watch?v=KzqpK1uCczw&t=790s
garrytan
garrytan @garrytan
Retweeted
Tim Ferriss Tim Ferriss
"The single most important thing for anybody wanting to break into any industry is go to the headquarters or cluster of that industry. Move to wherever that thing is. And all the advice that you can do anything from anywhere and everything's remote is all BS. With AI, 91 percent of private technology market cap is in the Bay Area. Ninety-one percent of the entire global set of AI market cap is all in one 10 by 10 area."
— Elad Gil
Listen to my interview with @eladgil:
https://tim.blog/2026/04/29/elad-gil/
garrytan
garrytan @garrytan
Retweeted
Silva Silva
Re @garrytan Gbrain deserves more recognition💯
garrytan
garrytan @garrytan
Retweeted
Gregor Zunic Gregor Zunic
With this going viral I forced myself to only use Codex CLI for last few days.
Took me a while to get used to the more autistic way of how codex speaks, but now it's good.
I admit it I was wrong. Codex fast + tmux + /goal is fucking insane.
@sama you guys cooked
Gregor Zunic: Who actually uses Codex over Claude Code? Claude Code is just 100x better imo, like the DX is WAY better.
sama
sama @sama
hey chat, we haven't forgotten about you

👀
GoogleLabs
GoogleLabs @GoogleLabs
Exciting Pomelli Updates!

👋New account! Go drop a follow and say hello to @PomelliByGoogle, your new home for all things Pomelli news and community!

🛍️ New feature! With Pomelli Catalog, you can sync your product or service details into Pomelli, helping you create personalized campaigns and images that match what you sell. Check out the post below to learn more! 👇

Pomelli By Google: Today, we're introducing Pomelli Catalog.📣✨

Add your products or services, and Pomelli will use them to generate personalized campaigns and high-quality photoshoots, tailored to your brand.

Free of charge. Available everywhere. Try it at: http://labs.google.com/pomelli

drfeifei
drfeifei @drfeifei
Retweeted
Astrocade Astrocade
We raised $56M to help build the next era of interactive entertainment. Series B led by @sequoia, Series A led by Sea.
Astrocade lets anyone create games with AI, play them with friends, and share them with millions.
But this isn’t about replacing creativity. It’s about giving more people the tool to bring their taste, humor, stories, and craft to life.
Today, the fun goes public.
sama
sama @sama
Retweeted
Michelle Pokrass Michelle Pokrass
we shipped gpt-5.5 instant today to chat; it's rolling out over the next couple days to everyone. for this model, we focused on factuality, crushing hacks, and improving the baseline intelligence. 5.5 is a pretty big step forward on all three. it is much smarter, significantly less likely to hallucinate, and just more delightful to talk to. there is always more to do but i'm pretty excited about this one!
drfeifei
drfeifei @drfeifei
So proud of you guys @_amirabs @sadeghian_ali ; it's been wonderful working with you guys, and seeing the progress of @PlayAstrocade !🚀

Astrocade: We raised $56M to help build the next era of interactive entertainment. Series B led by @sequoia, Series A led by Sea.

Astrocade lets anyone create games with AI, play them with friends, and share them with millions.

But this isn’t about replacing creativity. It’s about giving

garrytan
garrytan @garrytan
Retweeted
Paul Graham Paul Graham
Far-right trolls are tiresome, but they don't take up anything like the amount of attention that far-left trolls used to take up back in the 2010s. In those days it felt like a war. There were times when I only kept using Twitter to show that I couldn't be silenced.
AmandaAskell
AmandaAskell @AmandaAskell
"Wear a Claude-designed outfit to the met gala" is getting added to my list of life goals. Admittedly there are a few things higher on the list, but it's nice to add some fun ones.
sama
sama @sama
5.5 instant comes to ChatGPT today!

imo it is a pretty big upgrade, i really like using it.

Eric: Excited that we're updating the default model in ChatGPT today!

5.5 instant is a substantial improvement in intelligence, image perception, and factuality.

It also updates the writing style to be a bit plainer and more straightforward.

What was on your wishlist?
sama
sama @sama
i would like to talk to people who have built amazing things with 5.5 that weren't possible with earlier models. i am especially interested in examples that took ludicrous token budgets. thanks.
garrytan
garrytan @garrytan
Retweeted
orph orph
Alyssa Liu went viral during the Met Gala for her rant on the state of the public perception of data centers.
"We are being manipulated into ceding the compute frontier. Data center development is vital to national security." She adds, "Karen Hao's overstimation of data center water consumption by a factor of 1000 has done irreparable damage to the nation's efforts to stay ahead."
amasad
amasad @amasad
Retweeted
Michele Catasta Michele Catasta
Hundreds of thousands of you unleashed your creativity on Replit Agent this weekend -- completely free.
Our infrastructure handled ~4× the usual load, with tens of thousands of agents running in parallel.
A real test of our scale, and a powerful reminder of how passionate this community is.
Thank you for building with us 🧡
garrytan
garrytan @garrytan
Retweeted
Product Hunt 😸 Product Hunt 😸
Missed your YC shot?
We teamed up with @ycombinator: Schedule your launch on Product Hunt for this Friday and tag “YC application.”
@gustaf will review the top launches, and one or more will get a YC interview 👀
sama
sama @sama
the new instant model in chatgpt is so good damn

if you have been thinking-model-only for awhile, give it a try!
sama
sama @sama
in particular, the combination of improvements to speed, intelligence, personality, and great memory/personalization feels like a more-than-sum-of-the-parts thing when it all hits together

Sam Altman: the new instant model in chatgpt is so good damn

if you have been thinking-model-only for awhile, give it a try!
gdb
gdb @gdb
Major ChatGPT upgrade rolling out now, in the form of GPT-5.5 Instant:

OpenAI: GPT-5.5 Instant is starting to roll out in ChatGPT.

It’s a big upgrade, giving you smarter, clearer, and more personalized answers in a warmer, more natural tone.

And it's also more concise, which we heard you wanted. We think you'll love chatting with it.

garrytan
garrytan @garrytan
Come to the GStack GBrain Hackathon at YC

Y Combinator: We're hosting a GStack × GBrain Hackathon on May 16, with @TheHogAi, @transposevc, @askjoai, and @gbrainio!

Build and ship something real in 12 hours.

Win a YC interview and a 1:1 working session with @garrytan.

RSVP at http://events.ycombinator.com/GStack.

amasad
amasad @amasad
Half a million projects in a single day. One user used up $10k in agent workloads while another explored hundreds of business ideas.

Michele Catasta: Hundreds of thousands of you unleashed your creativity on Replit Agent this weekend -- completely free.

Our infrastructure handled ~4× the usual load, with tens of thousands of agents running in parallel.

A real test of our scale, and a powerful reminder of how passionate this

garrytan
garrytan @garrytan
Retweeted
Sudhir Mantena Sudhir Mantena
Re @garrytan I wish Gbrain setup was easier, but I understand these are early days. I wrote down a setup guide for first timers.
https://x.com/smantena/status/2051771339187613738
Sudhir Mantena: http://x.com/i/article/2051767458839822336
sama
sama @sama
5.5 in codex is so good for non-coding tasks.

i keep assuming it won't be able to do something, but a lot of the time i am pleasantly surprised.
ylecun
ylecun @ylecun
Retweeted
Mike Nellis Mike Nellis
Brutal montage from CNN of Trump lying about the ballroom being paid for by private donors
amasad
amasad @amasad
Retweeted
BetaBloom | Find your first users BetaBloom | Find your first users
Demo Studio now provides both portrait & landscape product demo from a single set of recordings.
* Auto-generated script
* No video editing skills needed
* Landscape & portrait from a single recording
Built with @Replit ✨
https://betabloom.app/demo-studio/
garrytan
garrytan @garrytan
Retweeted
Patrick OShaughnessy Patrick OShaughnessy
Brian sits on the board of Y Combinator. He said the last batch had 175 companies and only 16 of them weren't enterprise.
"Here are the reasons I think it's happening. Number one, when ChatGPT came out, people were afraid it was going to kill their business.
Number two, the business model is tricky. There is no consumer business model for AI that I've seen.
For example, ChatGPT, there's three ways it can monetize subscriptions. Unfortunately, they're probably going to hit a local maximum percentage of users.
Ads, they're hitting a local maximum because Claude and Gemini are not going to do ads.
And e-commerce, they shut down the third party apps.
And so the first thing is you need to have a business model around consumer AI. People are not trained to pay for information.
The second problem is distribution is mature. Like the app store. Now again, top three apps in the app Store are AI, so it does prove you have something revolutionary, you'll find your way to the top.
The third thing is, while I think Silicon Valley, we like to describe ourselves as rebels. I think it's very trend based and vibe based. And I think the trend is enterprise.
Maybe finally the reason people aren't doing consumer companies is that they're just harder. You have to be good at a lot more things. You generally have to be better at design, marketing, culture, press. It's not purely technology and sales.
But my prediction is that we're living in the age of enterprise AI, and I think in the next 12 to 24 months you're gonna see the beginning of a consumer AI renaissance.
Almost every app on my home screen has not changed since AI, including Airbnb. I think that's gonna change in two years."
Patrick OShaughnessy: My guest today is Brian Chesky (@bchesky), founder and CEO of Airbnb and one of the great consumer founders of the last 20 years.
Paul Graham coined "founder mode" based on Brian's experience running Airbnb. This conversation is about what comes after it, what he calls AI
garrytan
garrytan @garrytan
Retweeted
Yousef Yousef
We had @garrytan and @sdianahu come to give our lecture at Stanford today. Here is what I learned:
> “You can now boil the ocean” → previous expectations of what someone can do in front of a screen are outdated. A 6-person team can now hit $10M in revenue. In 2010, a 10-person startup felt impossibly lean. AI didn't just change the tools; it changed the range of outcomes completely.
> Evals are taste made executable → Generic benchmarks won't tell you if customers actually want your product. Founders are the ones who have to build the evals; nobody else can do it for you.
> Run your company as a closed loop → Open loop = information lives in human heads, decisions take days, errors compound silently. Closed loop = every workflow produces artifacts agents can read, every step has error-checking built in → now you get your 10x outcomes.
> AI is latent, code is deterministic → Latent is where intelligence lives. Deterministic is where trust lives. Build so you're not relying on the latent stuff when it matters.
> AI is changing the very structure of organizations → Middle management existed to route information between people. Closed-loop orgs don't need that anymore. Flatter teams, everyone shipping, and one person = one outcome.
YC's average batch is now hitting 10% week-over-week revenue growth. Five years ago that was the top 1% only. The world is changing. The question is whether you're changing with it?
And the even bigger question: will you be the one changing it?

YouTube

0

No recent videos fetched on this date.