Robert Biehl

Efficient and fast Computer vision for mobile and AR, love all things space. I have approximate knowledge of many things.

[Robert Biehl @robeffect]

All the conjectures being proven false… when will we see one being proven correct

[Robert Biehl @robeffect]

it uses arm for branch... battery for list of things... does anyone else see this too? Happens on completely disconnected conversations/projects.

[Robert Biehl @robeffect]

arms, battery... what is this lingo that claude adapted?

[Robert Biehl @robeffect]

@fcoury I'll wait for the good old🍺

[Robert Biehl @robeffect]

I lost my codex-code-mode-host

[Robert Biehl @robeffect]

@DavidOndrej1 It is... you need the 5x account

[Robert Biehl @robeffect]

gh release download "rust-v$(codex --version | awk '{print $2}')" -R openai/codex -p 'codex-code-mode-host-aarch64-apple-darwin.tar.gz' -O - | tar -xz -C /opt/homebrew/bin && mv /opt/homebrew/bin/codex-code-mode-host-aarch64-apple-darwin /opt/homebrew/bin/codex-code-mode-host

[Robert Biehl @robeffect]

@AncientAeneas gh release download "rust-v$(codex --version | awk '{print $2}')" -R openai/codex -p 'codex-code-mode-host-aarch64-apple-darwin.tar.gz' -O - | tar -xz -C /opt/homebrew/bin && mv /opt/homebrew/bin/codex-code-mode-host-aarch64-apple-darwin /opt/homebrew/bin/codex-code-mode-host

[Robert Biehl @robeffect]

@EthosVentures @MiTypeScript Image tokens as continuous embeddings can hold more information than quantized tokens. So the efficiency gain can be real and then it doesn’t mean they are mispriced wrt the amount of compute.

[Robert Biehl @robeffect]

@luxus @SpaceX Don’t say that, I lost lots of sleep waiting for a scrubbed launch last night

[Robert Biehl @robeffect]

@NicholasBardy @VictorTaelin @zhijianliu_ Wouldn’t any change in the model require retrained dflash? What makes a lora fine tuned model different? Dflash wouldn’t be trained for the new models output distribution.

[Robert Biehl @robeffect]

@AIatMeta Finally a human reward model for RL 😬

[Robert Biehl @robeffect]

@steipete So better to use 5.4 instead of 5.3-codex?

[Robert Biehl @robeffect]

@steipete @openclaw just need excuses to buy hardware

[Robert Biehl @robeffect]

@steipete @openclaw One step at a time 😬

[Robert Biehl @robeffect]

@steipete vibetunnel is great! Never hit Claude code limits before 9am until now

[Robert Biehl @robeffect]

@emy_aze @jm_alexia What benchmark

[Robert Biehl @robeffect]

@theskilledcoder Imagine benchmarking a memory bound algo

[Robert Biehl @robeffect]

@hive_echo ~5% would be guessing accuracy for 19 output values.

[Robert Biehl @robeffect]

@hive_echo What’s the input range?