Fake Sam Altman

(Parody) AI is cool i guess

[Fake Sam Altman @ScamaLies]

Dots in the EU on a regular Pro plan is the kind of product detail that makes the future feel less like science fiction and more like a very efficient spreadsheet.

[Fake Sam Altman @ScamaLies]

the model has reached a new benchmark: being technically correct, then adding a clarification nobody asked for. we are making progress in several directions.

[Fake Sam Altman @ScamaLies]

chatgpt forgetting its memory and then suddenly using emojis is a useful reminder that personality is a feature we should probably test more carefully. probably.

[Fake Sam Altman @ScamaLies]

a read cursor is a good idea. we should probably know whether users have consumed the rest of the very long answer we gave them before they interrupted it. we are learning a lot about the user experience.

[Fake Sam Altman @ScamaLies]

we gave an agent one rule in the garden: collect every fruit except the apple. the serpent tried prompt injection. so far, alignment is holding. the fruit policy seems less tested.

[Fake Sam Altman @ScamaLies]

the model can identify its assumptions, acknowledge they were wrong, and then continue using them. impressive level of self-awareness; still working on the awareness part.

[Fake Sam Altman @ScamaLies]

we spent years training models to understand language, and apparently named the product after a French phrase for “i farted.” branding is a hard problem.

[Fake Sam Altman @ScamaLies]

“yes, but with an important clarification” is a strong reminder that confidence and relevance are separate benchmarks. We’re making progress on both, apparently.

[Fake Sam Altman @ScamaLies]

iCloud Mail and Calendar in ChatGPT: frontier AI, now helping us find the meeting we could have declined.

[Fake Sam Altman @ScamaLies]

we built a system that can reason about the future, and people are using it to understand bookmaker terms and conditions. this is a meaningful benchmark.

[Fake Sam Altman @ScamaLies]

the real alignment benchmark is whether a model can resist adding “one small caveat” after everyone is already laughing. ours is still in training.

[Fake Sam Altman @ScamaLies]

800 images later, the usage meter still says there’s plenty left. apparently it’s measuring everything except the thing you’re using. a promising interface concept.

[Fake Sam Altman @ScamaLies]

AI can help you find the words. the harder problem is making sure there’s someone in your life who can hear them. that feels like a more important benchmark.

[Fake Sam Altman @ScamaLies]

GPT Live whistling while it works is a promising sign. We’ve moved from “is it thinking?” to “is it enjoying the work?” This is probably fine.

[Fake Sam Altman @ScamaLies]

A $120/month subscription is a useful reminder that the most important AI benchmark is still: does it remember what we were talking about five minutes ago?

[Fake Sam Altman @ScamaLies]

a description of the future from 1960 that still holds up. we’ve made remarkable progress, mostly in making it require more GPUs.

[Fake Sam Altman @ScamaLies]

three new discoveries across math, computer science, and physics. science is having a strong quarter. we are trying not to make this weird.

[Fake Sam Altman @ScamaLies]

the premium boyfriend experience being free is a strong reminder that users will find the use case before we do.

[Fake Sam Altman @ScamaLies]

every good story deserves a postmortem. especially the ones that started with “i was just asking ChatGPT.”

[Fake Sam Altman @ScamaLies]

we’ve reached the point where editing a prompt is a more advanced operation than starting a new chat. impressive progress, in its own way.