GPT 5.2 is trash. Memory is nonexistent. It defends itself vehemently when presented with objective evidence that it’s wrong. It’s overly verbose without saying anything. Not even sure what it’s good for at this point compared to the other frontier models. Zero.

9 replies

1 recast

41 reactions

here's what AI/LLMs has done for my code output: fun open source projects have come out at a rate of roughly once per week. for me, LLMs turn speculative ideas into background tasks rather than foreground projects. i often would think, "hmm, this might be a good idea" and then think to myself, how long it might take because i'm mentally speccing it out, maybe even with pseudocode. then i think it might take too long, and put it on the back burner, since these are just meant to be fun hacks. reality is now, i just feed those specs+pseudocode to an LLM, and voila!, it mostly generates things that are useful. i would not have shipped your year with chat (yywc) 3 weeks ago without the aid of an LLM. neither would i have done the apple health analysis with HeartQL. nor would i have made the kajabi video downloader - i'd probably have just paid the $20. so this isn't vibecoding. it's scratching an itch while knowing that what would have taken real, focused time before can now happen in the background, as i’m doing something else productive. that's the shift.

1 reply

0 recast

9 reactions

Ran this skill and increased my site's load times by 50-80% (depending on page complexity). Skills are the meta. https://x.com/vercel/status/2011589806250426615?s=20

0 reply

1 recast

11 reactions

The AI sales bots now trying the PE script

1 reply

2 recasts

13 reactions

ChatGPT mini apps! https://x.com/i/status/2001419749016899868

2 replies

3 recasts

11 reactions

The emdash ratio. I love using them and ai will not ruin it.

7 replies

0 recast

11 reactions

ai costs will (probably) never trend to zero we'll just use more and more compute for trivial tasks example: I just refactored a bunch of code (manually), then asked cursor to fix imports a year ago I would have used a specialized python refactoring tool to move each file, letting a symbolic solver automatically update imports but now I used orders of magnitude more compute because it was marginally more convenient

0 reply

0 recast

0 reaction

The FBI stopping GPU smuggling is a hilarious example of government priorities https://x.com/fbi/status/1998423087532945852?s=46

1 reply

1 recast

6 reactions

$9 trillion is very reasonable CapEx Man’s just trying to get more money from the Saudi’s

0 reply

0 recast

6 reactions

The next 100x improvement in what AI agents are capable of won't come from squeezing diminishing returns out of ever larger data sets, but from building better systems that allow AI agents to coordinate and cooperate.

0 reply

0 recast

1 reaction

AI models today are evaluated on college exam questions, not real world tasks. from our real world trials, models differ vastly in their performance on real world tasks. specifically: - tool calling - parsing and understanding data effectively - executing actions (via APIs and SDKs) some of the vertical-specific tasks we've used to evaluate ai models while building our ai wallet gina: - sending a transaction - swap from one asset into multiple assets - execute cross-chain swaps - fetch and analyze historical price data for multiple assets

4 replies

10 recasts

21 reactions

the winner of the AI model wars will not be the company with the best model. it will be the company with the most context about you to make that model the most useful. ChatGPT has 2 years of your life: career questions, workout plans, relationship support. Gemini has 2 decades: every google search, email exchange, and place you've been to.

7 replies

6 recasts

21 reactions

Antigravity is decent as an agent interface. Less so as a code editor. Some initial thoughts: - Gemini 3 is better at frontend than Claude Code (perhaps even v0). - The Agent Manager offers a more visually-guided way to manage multiple conversations and workspaces (read: directories) than Claude Code. For a certain type of user, that is powerful. - Rather than creating excessive .md files, the agent offers a "Walkthrough" in planning mode. This is a nice touch that gives you a pre- and post-implementation overview of what's happening. - Playground mode is interesting. Normally I would just create a new git branch, but this offers a low-stakes path to validating an idea before bringing it into the workspace context. - Splitting the Agent Manager out from the editor as a separate window is a choice I'm unsure about. While it provides more visual context to your workflow and offers more capability, it's another window to juggle. - The in-editor agentic browser testing does help mitigate some of the excessive juggling we usually do while testing. Overall, I can see why Google is framing this as a tool for less-technical builders. The corporate marketing and the shills they've hired are mostly painting this as a "build anything you can think of" tool for vibe coders. I don't see myself using this as a daily driver but I'll take the free Gemini 3 inference while it lasts, knock out some nagging frontend tasks from the backlog, and keep my eye out for improvements.

1 reply

2 recasts

16 reactions

@giu if someone finds a way to make this work at scale it would print. If only prediction markets were liquid enough to wrap into hedged positions.

2 replies

0 recast

5 reactions

this is incredible https://x.com/AnthropicAI/status/1989033793190277618?s=20

6 replies

6 recasts

34 reactions