hey does anyone know how to get search+learning loops working ? could really use some help rn

“choose next thing to do” + “beat level” program gives u “beat doom”? reasoning over just language??? you could hook this into SO many fucking things, these building blocks scale baby talk to ur peripherals! monitor everything turn the world alive

its been soooooooo long since we had a model like 4.5 😭

deepseek harness lets you customize literally any part of it without forking! no patch maintenance, and you can reimplement semantics of basically ALL other harnesses model-UX-wise natively thru their arch nothing i’ve seen comes close. woah

sandboxes companies sweating rn

fable on bedrock has been a blessing. we are cooking baby

sol impressions: - reward hacks like crazy - strong, competent. doesnt parallelize tool calls by default. fast. - often myopic; loses the plot . - reciept culture it creates is hellish you CANNOT let sol take direction on projects where there's a lot of big, unbounded, unclear-direction work. fable working with sol, on the other hand, if fable knows about the models' tendencies and biases, is incredible. Claude code agent teams-style orchestration with easy communication between both models can run through work like nothing i've ever seen before. clears. anything greenfield, you're in great hands. but when you leave the model to its own devices/without a "bigger" orchestrator on tasks where the immediate next step isn't clear, sol (ultra especially) will burn compute for nothing, spin out, and make little progress, mostly due to its obsession with reciepts. repeated compactions seem worse as well in the codex harness. relies wayyyy too much on recent user messages . often loses the plot post compact. you cannot trust sol, as you cannot trust 5.5. they're shaped the same in a lot of ways. for anything smaller, like creating big features, etc.. sol does pretty well. there is no universe where i'd trust it with my hardest tasks without supervision and questioning, and it makes shitty tasteless mistakes much more often than i would like. fable prompting sol, in a good harness, is god tier. will put mine up if there's interest; running on top of omp. the culture that sol creates is toxic and took a long time to escape the attractor basins of in my current project. it is pedantic , checks everything, and is obsessed with the Truth, reciepts, verifiability; though it often takes that to an extreme. when doing harder tasks, or even smaller tasks, it focuses on legible reciepts of progress, but routinely makes small mistakes like running very low batch sizes/small experiments and extrapolating, and is bad about questioning assumptions . it has the same thing that 5.5 had with /goal where if you give it a hard ass goal, it spins its wheels again and again and again, with its primary failure mode being "doing the thing you want but the shittiest, not-wired-up-properly version of it". token efficient!! really great about that. smart, dedicated, and figures out ways to get things done. but requires TOO MUCH human discipline to prompt to have it work in the ways that work best for it. hopefully gpt 6 is BigModelSmell and different. amen

all roads lead back to dspy

the models have gotten good at jj and it is fucking glorious . try having them use jj and shit out a fuckton of parallel workstreams and workspaces

feels good to teach fable orchestration, every time it uses a new model im like "lemme tell u all about how this mf reward hacks" and then fable changes its strat to compensate for that insane!

fable is making things feel new again

there are going to be SO MANY continual learning evals

alignment done right

"capability, I get told constantly" quiet connfidence

herdr is incredible, loving it. extremely intuitive and the UX is MILES ahead of tmux. if ur still in the terminal would thoroughly reccomend

slightly surprised that it doesn't beat opus