Tim Williams

@themishra4402 Sundar, quiet intelligence is underrated. The rest of these folks damage their reputation on a regular basis by opening their mouths.

If these four are building an ai startup together.... who should be the CEO ?

@elonmusk Probably want to build in space by that point.

Size does matter

@ThePrimeagen GitHub is feeling it in 2026

I cannot wait for the great data crisis of 2028

@h1kz0r @naval There's a lot more that can make it secure and deterministic, the terminal is a bad place to enforce controls on an LLM. Unless it's carefully restricted.

@naval AI is probabilistic, API access is what makes it safe, secure and deterministic. The terminal is a UI.

AIs replace UIs and APIs.

@Layton_Gott The only way they stop shipping slop is they work hard enough to cross the threshold into engineer. AI is way off from being able to create good, scalable, secure, unique products that have real value.

"Vibe coding" was SUPPOSED to be an insult. Yet it's becoming the most valuable skill in software... But only if you can actually do it well. Most vibe coders are shipping garbage. Half-tested features, made up APIs, and code that breaks the second a real user touches it. Learn to do it well or stop calling yourself a builder.

Here me out, github, but BroodWar themed.

@Zinny_Edmund Not for the attention it gets you, it matters if you actually leveled up through those commits, or just created slop PRs.

does github commits really matters ?

@_MattTrussell_ @justbyte_ A huge range of projects. I'm a team lead Software Dev full time, but since AI kicked off I've been knocking out a bunch of side projects as I go. Recently added a bunch of Open Source projects to my profile: github.com/tjameswilliams HBU?

@tim_willia98364 @justbyte_ What types of projects do you build?

As a vibe coder, what’s actually worth paying for right now? > $20 Claude > $20 Codex > $100 Codex > $200 Claude

@sama They're equally important, smarter SOTA for big tasks, cheaper (preferably open source, local models) for grunt work.

i keep thinking i want the models to be cheaper/faster more than i want them to be smarter but it seems that just being smarter is still the most important thing

@_MattTrussell_ @justbyte_ Cursor has improved a lot. Give it another shot, their house trained GLM 5.1 mode 'cursor 2' is decent and very cheap as well. I keep cursor around partially because I still need an IDE for complex things.

For sure. Complex code/integrations definitely trip up their logic. I was working on a rust project last year and it was hectic. The coding agents have improved a lot since then though. Every time I’ve tried cursor it’s been the worst performing of all the agents I’ve tried though. I tried the $20 plan for a few months. Won’t go back. I’ve had consistently better results with GitHub copilot.

@AradhyeAgarwal @sama You're switching to more expensive hardware. You pay a LOT more per token.

@sama What's the catch with the fast mode? What do you sacrifice to gain speed? Asking this since xhigh seems to already be specified, so it has to be something orthogonal to reasoning effort. Cannot be quantization also since then you would've said 5.5-mini.

5.5 xhigh in fast mode is really good i think i got psyoped by twitter on medium for a bit

@_MattTrussell_ @justbyte_ Exactly. I use Cursor, Claude Code and Codex daily. I bounce between them because I hate vendor lock-in. The 90% it gets right drops down to 70% when you dive into more complex domains. The less boilerplate your problem space is, the less helpful AI is.

@tim_willia98364 @justbyte_ Totally agree. I code with Claude 8+ hours a day, and I catch stuff all the time that I need to have it correct. It gets it right 90% of the time though. What are you building with Codex?

@elielAGI @justbyte_ This exactly. The way 'thinking like a swe' works now will change in the age of AI, but the skill of thinking procedurally the way coding trains your brain to do will become more important.

@tim_willia98364 @justbyte_ not a developer, but I think that in the current age syntax is irrelevant *but* thinking like a swe is useful and needed to properly leverage AI

@haider1 The more we try to understand about intelligence, the more we reveal how little we know. It's mass Dunning Krugers.

Sam Altman says a line from Ilya Sutskever that stuck with me: "prediction is very close to intelligence" If a system can compress the world into a smaller representation and predict what comes next, it starts to understand the data in a deep way That's the bet behind generative models

@icanvardar How will that happen when so much money is going into LLMs?

AGI running on consumer hardware is coming sooner than people think it’s not going to live in some distant data center forever

@sama Make me a hawk pet and I'll consider switching from Claude.

/hatch clippy

Kimi K2.6 tied GPT-5.5 on SWE-Bench Pro. Cost a fraction. Meanwhile JavaScript is finally fixing 30 years of date hell. Two massive open source wins. And I've never been more worried about open source's future. podcasts.apple.com/us/podcast/rub…

@Watershed423488 @justbyte_ Well that we agree on. How do you plan to solve the issue of 'shit amateurs' that you're likely not hiring now that would be your sr. engineers in 5 years? Or do you just believe LLMs will do it all?

Yes, I design now. Understanding what I still need people to do is part of the job. I need fewer people today than I did 6 months ago, and many fewer than I did 2 years ago. Even if we pretended ai coding skills scaled linearly, it's clear that shit amateur code is a short-term problem

@Watershed423488 @justbyte_ So you used to be a dev?

@tim_willia98364 @justbyte_ Thanks for trying. Maybe I should drop back down to dev for a project or two? It'll be nice to feel the superiority that comes with knowing "I typed all the characters for that". Will be sure to let you know if I do . Oh yeah, Dunning-Kruger...

@AndrewGinns Possibly also skewed by the fact that Anthropic has about half the amount of staff.

See loads of OAI peeps on X, never any A\. Is it just my bubble?