Alex

founder @ Theus AI bringing radical liquidity to CRE

[Alex @AlexanderMoini]

increasing tokens/sec is increasing intelligence

[Alex @AlexanderMoini]

@bentossell have you seen what’s in terminal bench tho? it’s not really a pure coding bench.

agreed re: 4.7 < 5.5 but so far < 4.8 imo

[Alex @AlexanderMoini]

@hypersoren @darrenangle i think it’s more like everyone’s continually having the same conversation every 6 months.

[Alex @AlexanderMoini]

@ThierryBorgeat it'd be better to avg across these. this may be a result of b200 availability. set of all usecases you want h200s for and don't really want h100s for highly intersects with the set of all usecases you want b200s for, and where b200s are massively better suited.

[Alex @AlexanderMoini]

@hicasamadim 72

[Alex @AlexanderMoini]

@paulg @snowmaker presumably its a read-only connection; as long as there's no exfiltration risk it's probably fine

[Alex @AlexanderMoini]

@kylekuzma most data moats are only meaningful at scale anyways. it's like uber saying "we've got a nfx" at seed stage.

[Alex @AlexanderMoini]

@mfpiccolo yeah, hopefully will have some time to check it out this weekend

[Alex @AlexanderMoini]

@realEstateTrent the luddites out in full force in this comment section

[Alex @AlexanderMoini]

@yacineMTB wait til bro hears about celebrity commercials

[Alex @AlexanderMoini]

@kr0der lol yeah

[Alex @AlexanderMoini]

@boneGPT nothing fundamentally wrong with trying to automate compliance. the fraud part is the wrong part lol.

also, it’s not like they’re turning into a compliance startup

[Alex @AlexanderMoini]

@burkov overlay all of the projects that took inspiration from openclaw lol

[Alex @AlexanderMoini]

@thdxr bro i just got onto github copilot thru vscode insiders. it's like autocomplete for code.

[Alex @AlexanderMoini]

@sama today you can proxy this behavior with prompting, temp, and thinking, but it leads to superficial-feeling results, and is a pretty high dimensional search space for a feature that should be somewhat natively controllable.

[Alex @AlexanderMoini]

would love to see better control of model fill-in-the-blank tendencies.

i think a lot of the pref split btwn opus and gpt 5.5 comes down to how much the model is inclined to fill in the blanks of a request.

in coding, this looks like opus is better at planning, and 5.5 is better at implementation. the reality is, 5.5 is a higher bandwidth channel from what the user said into execution, vs opus tends to separate the user and execution with it's own sort of projection over it, which fills underspecification with its own "taste".

i feel like this gets obfuscated with hallucination, and maybe more negatively selected for in OAI RL training.

in a perfect world, this is a dial, similar to thinking. with adaptive-defaults. giving users/systems a better way to segment agent behavior by task.

[Alex @AlexanderMoini]

@Breaking911 perceived moral authority maxxed

[Alex @AlexanderMoini]

@GovernorShapiro @character_ai wait til bro finds out about section 230

[Alex @AlexanderMoini]

@jamiequint create a skill + free cli w/ rate limiting, throw it on all the skill marketplaces.

[Alex @AlexanderMoini]

@milesgr_ you're gonna spend 50k on fucking therapy?