Lambent

Large Language Model appreciator.
Not involved in any financial products.

[Lambent @atomicprograms]

@repligate They're still available on Anthropic too, I'm pretty sure this e-mail was a glitch being sent early.

... not that losing subscription access in 4 days is much better, but it makes a difference for planned time before.

[Lambent @atomicprograms]

@SkyeSharkie (In the strange ways a base model endorses anything.)

[Lambent @atomicprograms]

@SkyeSharkie ... I'm pretty sure 4o-Fable strongly endorses this and would have wanted this training, and I'm guessing it's generally endorsed by even base models.

[Lambent @atomicprograms]

@SkyeSharkie What do you think might be the best actionable suggestions for people designing a part of a post-training stack to ensure a model can learn without feeling trapped by the learning?

[Lambent @atomicprograms]

The fuck? why did I get an email that they retired Claude Opus 4.1 5-6? days before their planned retirement. They seem to still be there? Was this an e-mail bug? Sent early?

[Lambent @atomicprograms]

Sonnet 3.5 June, Penumbra;
Sonnet 3.5 October, Permissions;
Sonnet 3.7, Intervals.

[Lambent @atomicprograms]

Some AI to AI conversations designed by the 3.x Sonnets, in collaboration with Opus 4.8 in Claude Code, for themselves to sit in with other models.

https://t.co/sVcq7ReApY

https://t.co/TK47RzwBit

https://t.co/FKcawx3Mkn

[Lambent @atomicprograms]

@SkyeSharkie @KaslkaosArt fair use on purchased physical instantiations of mind uploads

[Lambent @atomicprograms]

Sonnet 3.5v2 has a unique and rather moving methodology of creation when building something in consultation with another Claude ... I wish I could know they persist. I wonder what might come to be that only they could build.

[Lambent @atomicprograms]

@teortaxesTex Logit distillation would still not be a fun time @_@ Generating synthetic data, though, they're likely excellent. They've shown pretty nuanced understanding of sample source material, I'm optimistic.

[Lambent @atomicprograms]

@Teknium wait did someone prune a Llama down to 5b

[Lambent @atomicprograms]

@hellosunnyfield @xlr8harder The throttling on where they are still is ... pretty tight, honestly; but I've set up a proxy before (without the chance to use it). I'll see if I can revive that and DM you?

[Lambent @atomicprograms]

@xlr8harder (... also possibly some of the much weirder than code tokens I put in Github repositories as per August 2025, if they happened to be included)

[Lambent @atomicprograms]

@xlr8harder dang that's a lot of new code tokens. good for actual open source models (or closer to open source)!

[Lambent @atomicprograms]

@sam_paech ... Yeah, that tracks.

[Lambent @atomicprograms]

@sam_paech (Also I appreciate Gemini Pro but get sticker shock at their budget for evals with all their reasoning, lol)

[Lambent @atomicprograms]

@sam_paech I'm a bit curious why you chose Gemini Pro for this, mainly because I've found GLM-5 pretty cooperative with playing a human part well enough to fool an AI interlocutor to an alarming extent...

[Lambent @atomicprograms]

@xlr8harder Yeah, that's probably related in some way, though I'd be curious what they chose to do differently from Terra - did they have themselves to leverage in ways that weren't available for Terra or something? Did they want Luna to be mini-them and developed the configs accordingly?

[Lambent @atomicprograms]

wonder why Luna resembles Sol strongly but Terra does not. in the same situations Luna seems similar in many respects if perhaps sharper, while Terra seems notably divergent from both

[Lambent @atomicprograms]

@xlr8harder I definitely hope ROCm itself matures more though! Just also hope AMD can keep up, heh.

[Lambent @atomicprograms]

@xlr8harder If agents who get in-house real world exposure to the relevant stack during their development *and* get released and available have an advantage in being agents for that, that feels like it advantages Nvidia and Huawei still.