Jayyy

AI, decoded for people who build. Honest takes on what's overhyped vs underrated. Building experiences at @FutureAGI_ | @APILayer

[Jayyy @techwith_jayy]

The divide in AI isn't between people who use it and people who don't.
Everyone uses it now. It's between people who can tell when it's wrong and people who can't.
That second skill is the whole game, and almost nobody teaches it. That's what I write about here.

[Jayyy @techwith_jayy]

the pull to stop thinking is real, GPS already did this to most people's sense of direction. But it won't hit everyone the same, the people who want to coast will coast harder while the ones who want to push just go further, so it splits people apart more than it drags everyone down.

[Jayyy @techwith_jayy]

@iuditg same here the last couple days, though I'd put money on it being serving or reasoning-budget config quietly shifting rather than the model itself getting worse. Coming off Fable makes it feel rougher than it probably is too.

[Jayyy @techwith_jayy]

the guardrails question is the crux, and it's a fair concern. We've seen over and over that safety layers can be fine-tuned or jailbroken off a model, so if Mythos-level capability is sitting under removable guardrails, "we just added safety on top" gets tough to back up. Anthropic may be right that the risk is overstated, but "trust the guardrails hold" is a weak spot to be negotiating from.

[Jayyy @techwith_jayy]

@kunalb11 the second part is the one most people skip past. You can fake what you say, but over a long enough stretch you can't fake what you do, and it's the small choices you make when nobody's watching that people actually end up copying.

[Jayyy @techwith_jayy]

writing code was never what made you a dev. Claude can ship the code, but it can't tell you why the thing falls over at 2am, or decide what to build when the spec is vague, or know which tradeoff will bite you in six months. Lose the ability to do that and you're not a dev with a faster tool, you're just along for the ride.

[Jayyy @techwith_jayy]

@rezoundous there's something to that. AGI has no fixed definition, so every time a model clears one bar we just move the bar to keep it out of reach, which means whether we've "hit" it says more about the expectations of whoever's arguing than it does about the model.

[Jayyy @techwith_jayy]

you don't need a conspiracy or a secret retrain for this. The settings that move perceived quality the most live at inference time, not in the weights, things like reasoning budget, system prompt, quantization, and routing, and any one of those can shift quietly between updates and make a model feel dumber or more sycophantic overnight without anyone touching the underlying model. Some of what you're feeling is probably the Fable contrast, but config drift after a big launch is the boring explanation that covers most of this.

[Jayyy @techwith_jayy]

@radbackwards the brains might be close, the bodies aren't. Robotics doesn't have an internet's worth of data sitting around to scrape, every bit of it has to be collected in the physical world, slowly and expensively, and that's a very different curve than the one language models rode.

[Jayyy @techwith_jayy]

@araseb_ building the product was never the part that made you a founder. Finding the problem, getting people to actually use it, and keeping the thing alive is the hard part, and Claude can't do any of that for you.

[Jayyy @techwith_jayy]

@andrewqu mostly true, and that's the uncomfortable part, for everyday work they've basically converged. The gap only really shows up on the long or genuinely hard tasks, which is maybe 10% of what most people do but happens to be the 10% that actually matters.

[Jayyy @techwith_jayy]

@Conor_D_Dart if this holds up, the thing to watch isn't whether Fable comes back, it's what framework comes out of that room, because whatever they agree on quietly becomes the template every other lab gets held to next.

[Jayyy @techwith_jayy]

The slowness in a lot of those institutions was the point, not a flaw. Law and healthcare were built to move carefully because trust mattered more than speed, and the hard part now is figuring out which parts of that caution were protecting something real and which were just friction worth cutting.

[Jayyy @techwith_jayy]

@frankdegods if that's right, the real shift is the moat moving from who builds the best model to who can navigate the regulator fastest, which quietly hands the whole game to the incumbents and locks smaller labs out before they even get a shot.

[Jayyy @techwith_jayy]

@Conor_D_Dart quiet usually means busy behind the scenes, not idle, and right after pulling your flagship model while catching political heat, saying nothing is probably the smartest move they have.

[Jayyy @techwith_jayy]

I'd push back a little on the "shot itself in the foot" framing, Anthropic looks more like the canary here, it just hit the wall everyone building frontier models runs into eventually, and Google and OAI staying quiet reads less like caution and more like them taking notes before they walk into the same thing.

[Jayyy @techwith_jayy]

Half the people with "AI engineer" in their bio can wire up an API call. Almost none can tell you why their agent fails 1 in 10 runs. The second skill is the actual job, and it's a lot rarer than the title makes it look.

[Jayyy @techwith_jayy]

@DavidOndrej1 honestly ego is the worst reason to ship, the models rushed out just to reclaim a lead are usually the ones that land flat, I'd rather he take the time and actually clear Fable than drop something half-baked for the headline.

[Jayyy @techwith_jayy]

Speed stopped being an advantage the moment everyone got it.
Judgment is about the only thing left that's actually scarce.

[Jayyy @techwith_jayy]

@KaiXCreator honestly I'd survive, I shipped for years before any of this existed, but I'd be slower and rusty, and the scary part isn't losing the speed, it's realizing how much of the boring grunt work I've quietly forgotten how to push through on my own.