Daniel Hesslow retweetledi
GSPO is everywhere. It powers Qwen3. It "fixes" GRPO. But what is it?
We made a simple visual explanation. No jargon, no ML background needed. 👇
English
Daniel Hesslow
170 posts
Daniel Hesslow
@DanielHesslow
Making gpus go brrr in unison at @AdaptiveML
Katılım Kasım 2011
548 Takip Edilen288 Takipçiler
Daniel Hesslow retweetledi
AI agents aren't magic🪄🙅♂️. They're language models trained to predict very specific text.
When ChatGPT says 🔍"searching..." it's generating structured text that triggers actual tools.
Here's what's really happening 👇
English
Daniel Hesslow retweetledi
Some interesting use cases from @AdaptiveML powered by Gemma 🧵⬇️
English
Daniel Hesslow retweetledi
SK Telecom + @AdaptiveML trained Gemma 3 4B with PPO obtaining impressive results, specially for a model of such size
Learn more about how they did this adaptive-ml.com/post/sk-teleco…
English
Daniel Hesslow retweetledi
Kimi K2 is a vision into the future of 𝘀𝘆𝗻𝘁𝗵𝗲𝘁𝗶𝗰 𝘁𝗼𝗼𝗹 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻 and 𝗮𝗴𝗲𝗻𝘁𝗶𝗰 𝗺𝗼𝗱𝗲𝗹 𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴.
Leveraging 3,000+ MCP tools, the team generated 20,000+ synthetic tools and used them to train their 1T MoE model.
📜Paper and pipeline ⬇️
English
Shoutout from google is pretty great, nice work @alexchapeaux & and everyone over at @SKtelecom
Google AI Developers@googleaidevs
Build enterprise AI without the latency and cost of massive models. Learn how @AdaptiveML used Gemma 3 to create a multilingual customer service moderation LLM for @SKtelecom to support their 23M+ subscribers who speak a mix of English and Korean. deepmind.google/models/gemma/g…
English
Daniel Hesslow retweetledi
Build enterprise AI without the latency and cost of massive models. Learn how @AdaptiveML used Gemma 3 to create a multilingual customer service moderation LLM for @SKtelecom to support their 23M+ subscribers who speak a mix of English and Korean.
deepmind.google/models/gemma/g…
English
Daniel Hesslow retweetledi
Using Adaptive Engine, @SKtelecom tuned open models as small as Gemma 3 4B to exceed frontier performance (GPT-4.1, 3.7 Sonnet, and o4-mini) at multilingual content moderation.
Our research 📃 and full results 👇
English
Daniel Hesslow retweetledi
Looking forward to seeing everyone for ES-FoMo part three tomorrow! We'll be in East Exhibition Hall A (the big one), and we've got an exciting schedule of invited talks, orals, and posters planned for you tomorrow.
Let's meet some of our great speakers! 1/
English
We're back for the third year running!
ES-FoMo@ICML2025@ESFoMo
ES-FoMo is back for round three at #ICML2025! Join us in Vancouver on Saturday July 19 for a day dedicated to Efficient Systems for Foundation Models: from 💬reasoning models to🖼️scalable multimodality, 🧱efficient architectures, and more! Submissions due May 26! More below 👇
English
Turns out we have a twitter at Adaptive ML, who knew?
And it seems to be posting interesting stuff as well (completely unbiased observation)
Adaptive ML@AdaptiveML
Pretrained LLMs are aliens of extraordinary intelligence, yet little understanding. 👽 How do post-training techniques like 𝐒𝐅𝐓, 𝐑𝐄𝐈𝐍𝐅𝐎𝐑𝐂𝐄, and 𝐏𝐏𝐎 work in-tandem to turn these aliens into helpful AI assistants? 🧵 👇
English
@giffmana tbh it's not amazing, but it is less terrible than most other options. cargo is great & static typing let's you refacto stuff much faster. otoh the borrow checker really can be a pain for quick prototyping of stuff, and proc macros are a terrible hack that needs to be used a lot
English
Redpill me on rust. What should i read to give it a reasonable chance?
Knowing that i:
- 15y solid c++ experience up to including c++14
- think c++ is a bad language
- think adding types to Python was a mistake
- don’t foam on the mouth whenever i hear haskell
- am pragmat
English
@tri_dao Specdec improving throughput is a suuuper nice finding!
English
Surprisingly, speculative decoding works well not just for small batch LLM inference but also large batch and long context. Once we understood the compute & memory profile of LLM inference, the new spec dec algorithms fall out naturally
Together AI@togethercompute
We are excited to share our latest work on speculative decoding for high-throughput inference! Before this work, we thought speculative decoding was useless at large batch sizes since the GPUs would go brrrr from processing all the different inputs. Much to our surprise, we discovered speculative decoding is quite useful if the inputs are long enough because decoding once again becomes memory-bound from the large KV cache. In fact, we show that speculative decoding can improve latency AND throughput by up to 2x in this regime! Read more here: together.ai/blog/speculati…
English
@Muennighoff @zach_nussbaum I was a bit unclear, information can leak in the forward pass from future tokens into the previous ones. Here's an illustration, you can completely predict the next token, through leakage in the routing.
Obviously very toy example, a single expert that choses a single token
English
Releasing OLMoE - the first good Mixture-of-Experts LLM that's 100% open-source
- 1B active, 7B total params for 5T tokens
- Best small LLM & matches more costly ones like Gemma, Llama
- Open Model/Data/Code/Logs + lots of analysis & experiments
📜arxiv.org/abs/2409.02060
🧵1/9
English
@Muennighoff @zach_nussbaum (And assigning it to the last "hello" is just a matter of counting the number of previous hellos, and have the routing be a function of that).
I guess maybe you can't actually leak that much information through this. Might just be too costly to be worth exploiting it
English
@Muennighoff @zach_nussbaum Like you should be able to have one expert that says there's no more "hello" and assign it to the last "hello" in the sequence or smth. But y'know sometimes the optimization process is not strong enough to be able to exploit every loop hole. Cool finding tho!
English
@Muennighoff Interesting about EC vs TC, how do you do expert choice with a causal model?
English
OLMoE Experiments
1) Expert granularity i.e. we use 64 small experts per layer
2) Dropless token choice beats expert choice routing
3) Shared experts worse for us
4) Sparse upcycling not helpful in our regime
Check the paper for more🙂
🧵4/9
English
This is a very nice direction from @PyTorch! Even when we need the highest possible performance, we can still use Torch as a first step and export the IR to external codebases with production guarantees around memory etc!
English
You can even get the compute graphs for training, which is super useful!
English
Want to know what actually goes on inside a PyTorch function? Found a new undocumented feature that shows it 👀
English