Max @Maxmatical

llm training Joined September 2013

Tweets

480
Followers

54
Following

654
Likes

2K

Max @Maxmatical

a month ago

@steven_lebron Thats my left leaning librarian goat

0 0 0 8 0

View Details

Max @Maxmatical

a month ago

@PandaAshwinee why do you say pretraining specifically? seems like one of the main axis of improvement is over long time horiztons/inference time scaling, and that's mainly a product of rl no?

0 0 0 89 0

View Details

Cohere @cohere

a month ago

Introducing Cohere's first open-source coding model: North Mini Code Small & efficient, designed for agentic performance and built for community input.

69 266 2K 594K 1K

View Details

Cohere @cohere

4 months ago

Introducing: Cohere Transcribe – a new state-of-the-art in open source speech recognition.

81 293 3K 612K 2K

View Details

Max @Maxmatical

4 months ago

@natolambert @thdxr Dont forget the og starcoder(2)!

0 0 0 317 0

View Details

Max @Maxmatical

5 months ago

rip to the ai labs that are just qwen finetunes

0 0 1 39 0

View Details

Max @Maxmatical

9 months ago

@eliebakouch @LiamFedus @barret_zoph fast-llm from servicenow research

0 0 1 225 0

View Details

Max @Maxmatical

9 months ago

i keep trying to use := in my code and it never works out well for me

0 0 0 24 0

View Details

Max @Maxmatical

9 months ago

another win for @ServiceNowRSRCH

This is the most impressive plot I've seen all year: - Scaling RL not only works, but can be predicted from experiments run with 1/2 the target compute - PipelineRL crushes conventional RL pipelines in terms of compute efficiency - Many small details matter for stability &

11 22 271 30K 179

0 0 0 75 0

View Details

Max @Maxmatical

11 months ago

@corbtt @winglian @jayendra_ram So youre comparing ood perf vs training on in distribution data and claiming frontier performance? Oof

0 0 0 18 0

View Details

Max @Maxmatical

11 months ago

@winglian Training on test remains undefeated

0 0 1 73 0

View Details

Max @Maxmatical

a year ago

@xeophon Remember to give reka $$$ if you want more models kthx

1 0 0 118 0

View Details

Reka @RekaAILabs

a year ago

Reka Research is our AI agent that scours the web to answer your toughest questions. Ready to unlock its full potential? Learn directly from the team who built it!

2 13 46 7K 13

View Details

Max @Maxmatical

a year ago

weekly reminder that you should be trying reka flash and reka research

0 0 1 40 0

View Details

Sharath Raparthy @sharathraparthy

a year ago

looking for an agent which excels at questions that require dozens of sources and delivers accurate responses with reasoning traces in a few minutes? Reka Research is here for you. start building today: docs.reka.ai/quick-start

0 5 18 1K 2

View Details

Max @Maxmatical

a year ago

@teknium how so? this just shows the training loss, which could be easily explained by the lr schedule. you've already seen this type of loss in deepseek v1 with their lr schedule

0 0 7 1K 1

View Details

Max @Maxmatical

a year ago

Your daily reminder to try reka flash 3.1 and reka research

0 0 0 44 0

View Details

𝚐𝔪𝟾𝚡𝚡𝟾 @gm8xx8

a year ago

Reka Flash 3.1: an upgraded coding-optimized 21B LLM with strong agent finetuning potential - +10pt on LiveCodeBench v5 (Full) vs Flash 3 → from improved RL training with verifiable rewards (RLOO) - Uses a REINFORCE variant with token-level loss, on-policy updates, DAPO-style long-sample handling, and dynamic sampling - Converts multi-choice math questions to fill-in-the-blank to avoid reward hacking - RL filters low-quality math examples; code rollouts execute trajectories on the fly - Post-trained on public + synthetic SFT data, then RL-aligned on math/code (Numina-1.5 + executable test cases) - Competitive with Qwen3-32B, o3-mini, Gemini 2.5 Flash on code tasks - Powers Reka Research, an agentic system for document/web QA - Llama-format + 3.5-bit quantized version released for local use - Primarily English; limited multilingual ability

1 13 57 3K 24

View Details

Florian Brand @xeophon

a year ago

I really enjoyed @RekaAILabs Flash 3, people overlooked that one. New model still uses REINFORCE, but with DAPO-like modifications. Other new(?) thing: Concert MC -> FIM to lessen reward hacking