Yeah, I've avoided most of that stuff too. We do have an Echo, a Google Home speaker, and some Sonos speakers that are kind-of IoT I guess... Oh, and a couple of smart plugs so I can turn on some lamps remotely or on a schedule.
"Claude is Learning to Design..." " ... train Claude specifically on chip design and verification, teaching it ..." This characterization of Claude as a model is very common, but I've never understood it. Claude is a harness that runs models such as Opus and Sonnet that can learn and be trained, right? In this sense it's like ChatGPT, but I don't think I've seen the same usage treating ChatGPT as a learner and trainee.(Both Claude and ChatGPT can start and manage agent software that in turn uses models, but I don't think that's relevant to my point.)
The way I understand it, Claude is a family of models. Anthropic trains very last base model, then post-trains them with tons of RL and such. Once they have something they like, they cut them down to sizes they can serve (quantization, distillation, having larger models post-train smaller ones, etc), and that's where the specific models come from (Haiku, Sonnet, Opus, Fable/Mythos).
The equivalent for OpenAI is the GPT-5 family or the GPT-6 family, which has different sizes (Luna, Terra, Sol, Astra).
Then these models are put in harnesses that steer then and help them manage context, multile agents, how to deal with local computer use, browsing, etc (and that's Claude Code, Cowork, Codex, ChatGPT Work, etc).
The labs have internal models that are both generations ahead of what customers have access to, checkpoints along bigger training runs, or experiments, or models with fewer guardrails, etc. So whatever they do internally probably isn't "Opus" or "Astra" necessarily working on the R&D stuff.
That's my understanding as an outsider, so I could be wrong. Thanks for the comment, Steve 💚 🥃
I think you’re going to love that film and I’m excited to see the recaps here or in Discord!
I did really enjoy it! I posted some thoughts here: https://www.libertyrpf.com/i/223326965/the-arts-and-history
This is what I get for being an issue late in my read list haha thanks
No worries! Thanks for reading, I appreciate any time you give to my scribblings 💚 🥃
I never got a Ring type camera or an internet connected refirgerator because I have a healthy fear of the IoT. Call me a Luddite.
Yeah, I've avoided most of that stuff too. We do have an Echo, a Google Home speaker, and some Sonos speakers that are kind-of IoT I guess... Oh, and a couple of smart plugs so I can turn on some lamps remotely or on a schedule.
"Claude is Learning to Design..." " ... train Claude specifically on chip design and verification, teaching it ..." This characterization of Claude as a model is very common, but I've never understood it. Claude is a harness that runs models such as Opus and Sonnet that can learn and be trained, right? In this sense it's like ChatGPT, but I don't think I've seen the same usage treating ChatGPT as a learner and trainee.(Both Claude and ChatGPT can start and manage agent software that in turn uses models, but I don't think that's relevant to my point.)
The way I understand it, Claude is a family of models. Anthropic trains very last base model, then post-trains them with tons of RL and such. Once they have something they like, they cut them down to sizes they can serve (quantization, distillation, having larger models post-train smaller ones, etc), and that's where the specific models come from (Haiku, Sonnet, Opus, Fable/Mythos).
The equivalent for OpenAI is the GPT-5 family or the GPT-6 family, which has different sizes (Luna, Terra, Sol, Astra).
Then these models are put in harnesses that steer then and help them manage context, multile agents, how to deal with local computer use, browsing, etc (and that's Claude Code, Cowork, Codex, ChatGPT Work, etc).
The labs have internal models that are both generations ahead of what customers have access to, checkpoints along bigger training runs, or experiments, or models with fewer guardrails, etc. So whatever they do internally probably isn't "Opus" or "Astra" necessarily working on the R&D stuff.
That's my understanding as an outsider, so I could be wrong. Thanks for the comment, Steve 💚 🥃