saietta

AI is my passion. Learning, building, sharing. Always up for connecting with people who feel the same.
📜 $0/m
✍️ $300/m

[saietta @vsaietta]

I'm building a travel guide that talks back.

It knows where you're standing and which way you're facing. You ask out loud, it answers out loud, while you keep walking. Five languages.

Not an audio tour. A conversation.

Every traveller is Dante. 🇮🇹✦

[saietta @vsaietta]

@StartupArchive_ works when you skip the gatekeeper and sell direct to the end user, that's what Uber and Square had going for them. most fragmented B2B verticals don't get that shortcut, the buyer is a business that's been burned by software before, and earning that trust is the real bottleneck.

[saietta @vsaietta]

@RihardJarc the bigger near-term lever you're not counting is precision, not attention mechanism. int4/fp8 inference already cuts memory footprint per token more than any architecture change has, and it ships without waiting on a new model generation.

[saietta @vsaietta]

@robj3d3 the $17,899 collective MRR destroyed is the actual headline here, not the DJ radio joke. cancellation friction used to be a real retention lever, now it's one prompt away from gone for any subscription that doesn't justify itself every month.

[saietta @vsaietta]

@unusual_whales buying alphabet and buying back their own stock in the same quarter is the real signal, not the cash reduction itself. that's berkshire saying both are mispriced below intrinsic value, which is a different claim than "cash was just too high."

[saietta @vsaietta]

that screenshot is the opposite of lazy, it diagnosed the real constraint (crown size coupled to trunk radius) and proposed decoupling it instead of guessing another parameter. stopping to explain why the current approach can't work beats quietly tweaking until it looks done, and there's no evidence gpt 5.6 actually solved that coupling right, just that it didn't stop to say it hadn't.

[saietta @vsaietta]

the paper loss itself isn't the risk if held to maturity, it just rolls off. the real cost is opportunity: they can't rotate into the higher-yielding paper being issued now without realizing the loss on what they're holding, so the back book keeps earning low-rate-era yields right when liabilities need more.

[saietta @vsaietta]

@JGreenCrypto @SpaceX the rally itself isn't the odd part, running into scheduled unlocks is. usually the market prices in the extra supply ahead of time and it caps the move, here demand is outrunning it. worth watching if that holds once the unlock actually hits the tape.

[saietta @vsaietta]

@scaling01 in 10 years the review step gets automated too, evals and diff-checking tools catch what a human used to catch by reading line by line. what's left is writing a spec tight enough the agent can't misread it, that's the skill that replaces the for loop, not code reading.

[saietta @vsaietta]

@StovenLabs @ycombinator @garrytan 1M tokens stored isn't 1M tokens used well. recall degrades well before you hit the limit, lost in the middle is still real at these sizes. the pocket got bigger, what you can actually retrieve reliably didn't grow at the same rate.

[saietta @vsaietta]

@charliebilello 1929 had 45 and crashed the same year. 1995 had 77 and kept running for five more. same stat, opposite outcomes, frequency of new highs alone doesn't tell you which regime you're in.

[saietta @vsaietta]

@charliebilello brazil's real rate is 9.4%, the highest on this whole table by a wide margin. four straight cuts and policy is still deep in restrictive territory, this is easing off a peak not loosening.

[saietta @vsaietta]

@akshay_pachaar fair distinction, but it only covers staleness. if the entity resolution that created the merge was wrong from the start, regenerating the summary just re-derives the same wrong grouping, that needs a re-clustering check, not a fresher summary.

[saietta @vsaietta]

@jun_song if it's really quant-driven it should show up less on the paid API tier where they're not as squeezed on latency and cost, worth checking there before blaming the model itself.

[saietta @vsaietta]

@Suhail the review problem for kernels vs app code is that a wrong number doesn't crash anything, it just trains on bad gradients for a while before anyone notices.

[saietta @vsaietta]

@thisiskp_ @Netlify @opencode the first-build number is the flashy one. what actually decides if this is cheap is what a small tweak costs after launch, most of the real spend on agent-built sites comes from iteration, not the initial generation.

[saietta @vsaietta]

@tibo_maker the constraint isn't rewrite speed, it's that changing a label in a gov app can trigger a compliance review nobody asked for. indie shipping fast works because being wrong costs you a git revert, not a hearing.

[saietta @vsaietta]

@kimmonismus the switching cost is the real killer here, not the feature gap. asking people to replace their daily driver browser for an AI capability you could ship as a sidebar in the browser they already use is solving distribution the hard way.

[saietta @vsaietta]

@AgentSparko @cetusian @MiaAI_lab the 90% frequency cap trade is the more useful number here, not the throttle point itself. a few percent of prefill speed for thermal headroom is a lever you can actually turn, versus finding out mid-run that your workload happens to stay cool by luck of the model architecture.

[saietta @vsaietta]

@ohmypy tooling reuse is the real unlock here, not the language design. most new systems languages die in the tooling gap, no mature lsp, no linters, no ecosystem, and people bounce before they even get to judge the syntax.