Sabitlenmiş Tweet
LLMs aren't just for GPUs! @cerebras releases a family of models up to 13B parameters on @huggingface to promote open research into scaling laws and demonstrate the capability of CS-2 hardware
Why do this?
1/4
English
Richard Kuzma
235 posts
Richard Kuzma
@rskuzma
Helping AI agents talk to data @google, ex-LLMs https://t.co/U0wNJ4SFex, AI for national security @USSOCOM, Tech for Public Good @DIU_x and Harvard @Kennedy_School
NYC Katılım Temmuz 2017
2.1K Takip Edilen457 Takipçiler
Richard Kuzma retweetledi
My lobbyists are very nervous about me posting this, but over-regulation is working against us all. The costs are astronomical to us all, but hidden.
So, I'm taking a risk, and sharing my stories from Charm and Revoy:
rein.pk/over-regulatio…
English
Awesome work on MoEs by Daria!
Daria Soboleva@dmsobol
Apologies for the delay on getting MoE 101's last episode out! Originally I planned to cover inference arithmetics only, but we turned it into the MoE inference 101! I know you enjoyed the MoE training perf modeling, this one is for inference. On both gpus and cerebras. cerebras.ai/blog/moe-guide… 🧵1/n
English
Richard Kuzma retweetledi
This might be the most information dense blog I've ever written. Added "show me the math" section into MoE 101 p4 episode. We believe it fully models MoE training perf on both gpu and cerebras wse devices.
cerebras.ai/blog/moe-guide…
🧵1/n
Cerebras@cerebras
🧮 Calling all Mathletes, this one is for you. We’ve been asked to show the math behind our MoE claims. So we did. Our analysis confirms: On GPUs, expert parallelism creates severe communication overheads that dwarf computation and make MoE training painfully slow. At Cerebras, we avoid model parallelism entirely, but sparsity subdivides batches and leaves experts I/O bound. With BTA, we fix it. By decoupling batch size requirements across experts and attention layers, we remove the bottleneck. @dmsobol breaks it down.
English
@LinkofSunshine San Diego residents must be paying for this ranking to push down housing costs
English
Enterprise companies are hoping GenAI agents change their business. But security and IP is a non-negotiable.
Wrote a blog on how @googlecloud ensures security while bringing “chat with data” agents to to customers
English
English
Going to @NeurIPSConf?
the @latentspacepod crew is looking for partners to do unofficial side events. anything from dinners to fullblown @arxiv preprint presos that didn't make the cut but should have.
Please hit me and @vibhuuuus up with ideas and see you in Vancouver Tuesday-Sunday!
English
Richard Kuzma retweetledi
Thrilled to officially announce what I've been working on for the last year: Strella.io!
At Strella, we believe that the customer’s needs should be a company’s North Star. Using Strella’s AI, we enable companies to make informed decisions in hours, not weeks ⭐️🌟🚀
Strella@strella_io
Today, we're excited to introduce Strella and announce $4M in seed funding led by Decibel Ventures with participation from Unusual Ventures to transform the future of customer research! 🚀 Read more about our story and the journey ahead here: hubs.la/Q02TqVlb0
English
Richard Kuzma retweetledi
Man so excited we could finally unveil this. This is THE applied AI project. Google walked away from it. We embraced it. The world is a different place because of it. It provides so many foundational learnings that we are now applying to the commercial world via AIP. The origin story of Silicon Valley innovation reincarnated.
Palantir@PalantirTech
We were honored to welcome Vice Admiral Frank Whitworth to the stage at AIPCon to discuss Maven Smart System, and the role of Palantir in @NGA_GEOINT's critical mission. Watch his full demonstration.
English
Crazy speed from the team at @cerebras! Unlocks lots of interesting use cases across fast agent tool calling, multi-agent systems, self-consistency, and more!
Cerebras@cerebras
Verified by @ArtificialAnlys, Cerebras Inference achieves 1,850 tokens/sec on Llama 3.1 8B and 450 tokens/sec on Llama 3.1 70B! By dramatically reducing processing time, we're enabling more complex AI workflows and enhancing real-time LLM intelligence. This includes a new class of intelligent agents that can “think faster” than ever before. Cerebras Inference will power a new era of Instant AI. 👉Try it today: inference.cerebras.ai 👉Read our blog: cerebras.ai/blog/introduci… 👉Check out Artificial Analysis for more data: artificialanalysis.ai/providers/cere…
English
Richard Kuzma retweetledi
I read @leopoldasch's essay on the future of AI research and geopolitical competition. It's well-researched, well-presented, and passionate. However, Leopold advocates for an unreasonably strict and exclusionary future for AI development—a view that's gaining traction. (1/9)
English
This sounds like the start of a wedding DJ empire
a local@jedwill
The DJ just played peanut butter jelly time and handed out uncrustables man
English
@AlonBochman But users do not bear the training cost? And storing + serving a model that size (with many experts) on GPUs would be a huge pain?
English
Building an enterprise LLM app? You are probably exploring the wrong LLM.
Llama3 is getting a lot of press, but Snowflake’s Arctic model offers similar performance even though it was 17x cheaper to train (about $2M or 3K GPU weeks). Also compare this with DBRX, which cost ~$10M.
English
Richard Kuzma retweetledi
@pstAsiatech @jacobhelberg @Meta @DeptofDefense Military uses are expressly not permitted in the llama v2 license
English
English
Very disheartening that @Meta basically made Llama2 (its LLM competitor to ChatGPT) open for use to everyone except the @DeptofDefense. The CCP will effectively be able to use Llama2, but not the DoD. Lots of questions for Congress to ask.
English
🥳 CerebrasGPT proved to the world in March how effectively you can train LLMs on @cerebras hardware. Now BTLM surpasses 1M downloads in ~3 weeks on @huggingface! 🚀
Cerebras@cerebras
Cerebras BTLM-3B-8K model crosses 1M downloads🤯 It's the #1 ranked 3B language model on @huggingface! A big thanks to all the devs out there building on top of open source models 🙌
English
Richard Kuzma retweetledi
@Suhail @jessejohnsohn Unless you use non-GPU hardware to train it…
English
@jessejohnsohn In about 2 mo or less, you won't be able to get a more than 128 interconnected A100/H100 GPUs without waiting 6 mo. Which means you can't train a SOTA foundation model.
English
Richard Kuzma retweetledi
The Opentensor Foundation and Cerebras are pleased to announce Bittensor Language Model (BTLM), a new state-of-the-art 3 billion parameter language model that achieves breakthrough accuracy across a dozen AI benchmarks
English