Devoted to addressing alignment. We develop state of the art open sourced AI. https://xitter.cf/t.co/oANsMnut7V https://xitter.cf/t.co/6aJDLUvuU5
Your Digital Ecosystem
Joined April 2023
- Tweets7.2K
- Following4K
- Followers12.5K
- Likes1.7K
you know there really is a lot more going on there thats practically useful than it seems like at face value
still blows my mind arxiv.org/abs/2403.17887
ALT layer similarity metrics in a modern llm showing most layerss are redundant and can safely be removed
not bad for a byte level model in 1 hour on a single local gpu from scratch, especially one under 45m params B)
ALT exceptionally performant byte level AI achieving 73% test accuracy on fineweb edu in a very short period of time on a single consumer gpu
@huggingface was it intentional to put a paywall on the search bar of huggingface.co/docs ?
oh this makes so much sense, no wonder its been impossible to get to write any code lately, what are you guys back to 512 token windows again?
i thought everyone was on the same page that this is an ineffective and not useful strategy for reducing context issues
i cant believe you guys are stuffing the things context window with my past chats and enormous system prompts but then truncating the actual conversation
what are you building? its not something designed to be a *use*ful product...
ALT a screenshot of openais 'thinking' panel showing that GPT-5 pro is not using the conversations history for inference, but is insteead consuliung a summary_reader for summaries of the previous turns.
there should be an option to always allow messages from followers, choosing between bots or filtering low quality messages means that either way i risk missing reach outs, and the entire purpose of this account is to be accessible to people puiblicaly
ALT the direct message menu on X with settings
@AnthropicAI
im paying 100$ a month, i sent one single prompt in the last 3 days , this is one message into the chat
how big is this thing? have you tried tokenformers?
has anyone else noticed the frontier models are 'hallucinating' that the user is talking to them behind the scenes?
seems interesting to me
and a little creepy, but those that know me well have heard me say more than once that this is the only way to progress beyond some point
youtube.com/watch?v=oA9qTxqJ…
snipping out clips from the podcast! the open source is all weve got folks
@ToolUseAI
alignmentlab.ai
just in case you missed it, we did a podcast with mike @ToolUseAI about our new ai optimized workstation, how we plan on helping the tech align to the people and what we expect to see going forward, give it a watch and give tool use a follow!
alignmentlab.ai
check us out on @ToolUseAI with @MikeBirdTech
talking about Senter , science, and alignment!
reach out to us and get Senter, and have your own personal AI workstations at our discord:
discord.gg/TmSR5unjek
and pop in to chat later tonight on the tool use discord!
discord.gg/7x6KeZ54JF