jlf

[jlf @jlffinance]

@Dorialexander At this point it the models just need to be more strategically intent-aligned to be superior in finace they rest they get. is it like super narrow detailed transaction traces?

[jlf @jlffinance]

@TheZvi @umnovd So does this leave us with people maybe hedging themselves elsewhere in correlated bets or actually a market that provides us with some reliable pointer? Also given the volatility alone of these contracts, I‘d not take them too serious(?)

[jlf @jlffinance]

@umnovd @TheZvi What‘s your read? Do you know more about prediction markets and how these move among low liquidity? I.e. if I put 5k in a yes until 31 Jund rn, how would that move the market?

[jlf @jlffinance]

@ClaudeCodeLog fix the context with opus

[jlf @jlffinance]

@blueblimpms Meaning whether you‘ve seen anyone do a good write-down on this*

[jlf @jlffinance]

@blueblimpms Hey man, essentially checking this real-time too. Have you seen what the implications of such intervenations are on the financial markets/ ai trade anywhere? Because this could become a forcing factor if it‘s bearish (as GDP-growth tied closely to the compute build-outs).

[jlf @jlffinance]

@DanielCHTan97 i noticed the same thing especiall yesterday, it‘s a blend of fable / opus - much more opinionated & just working through tasks. I recently adjusted my claude.md & deactivated some hooks which might explain some uplift, but not so much…

[jlf @jlffinance]

@argofowl medium until i figure out a way to compile context packages perfectly per task so it one shots and i don‘t need long ass threads

[jlf @jlffinance]

@KingBootoshi cool, me too - wanna chat from solo to solo? I kinda leverage these systems like you do (maybe a bit less looped ;) ), but in another domain

[jlf @jlffinance]

@KingBootoshi okay good, if you mind me asking, are you working for yourself or are you employed? and how do you measure the (latent) value of what you‘re building?

[jlf @jlffinance]

@JJ_StuJordan @paulbettner @TheRealAdamG @AnthropicAI @OpenAI opus has gotten a lot worse for this architectural stuff withe 4.8 & therefore has lost it‘s edge for me

[jlf @jlffinance]

@infohazard_lol i agree - for me opus 4-7 was for maybe a few days only the best model re orchestration and intent-aligning - it could spit out 400-1000 lines in one write which made it the pefect context-setting system for sub-systems or agents. now it does 50lines for same set-up. useless.

[jlf @jlffinance]

@doodlestein is that even allowed? diff accounts on same machine?

[jlf @jlffinance]

@thsottiaux When using in Cursor terminal as CLI via windows sometimes the answer gets truncated and just stops mid sentence

[jlf @jlffinance]

@KingBootoshi i wanna feed this to my system so you got code/redacted session transcript chunks or so?

[jlf @jlffinance]

@yoheinakajima i chunk my cc/codex sessions and use them vis an agentic retrieval harness as decision traces - any kind of framework will blow as soon as the next model comes out trust me or writing to md/html files will just bloat + bias the model towards a weird attractor basin it‘s state.

[jlf @jlffinance]

@doodlestein @VictorTaelin plus i build a dispatch tool where it can outsource taks to gpt 5.4 xhigh fast via the sdk for obvious codex work (properties of <good code> are clear)

[jlf @jlffinance]

@doodlestein @VictorTaelin i build a script with hooks to enforce a) exploration-depth deterministally (hook will fire and tell system it needs to do X more uses of the read / any own retrieval tool i built) and b) enforce FULL file reads not just lines=50 or head -50 or so) - works like magic