Hari Subbaraj retweetledi
Hari Subbaraj
38 posts
my fav part of this is @be4ncurd writing a weekly meta automation that scans all triages, updates its own investigation skills, and then tells us what components need a deeper look. @DevinAI now my @usepylon for internal system complaints 🙌🏾🐐🫡
Cognition@cognition
Introducing Devin Auto-Triage: Your AI first-responder with long-term memory. Devin can monitor incoming bugs, alerts, and incidents, investigate them, and come back with context, next steps, or a PR.
English
Hari Subbaraj retweetledi
Hari Subbaraj retweetledi
Throughput vs. latency is a foundational tradeoff in almost any system. @sailresearchco is building inference for long-horizon agents, so we’re all about the former - exactly as Ben predicts! Software is our first lever here, and in time we’ll rebuild the whole computing stack.
Stratechery@stratechery
The Inference Shift Agentic inference is going to be different than the inference we use today, and it will change compute infrastructure because speed won't matter when humans aren't involved. stratechery.com/2026/the-infer…
English
Exciting to see what @andrew_li03 and co are cooking up!!!
Alex Shan@alexshander03
We’re launching @JudgmentLabs today and announcing $32M in funding. As AI agents take on more of the work that creates economic value, they generate massive amounts of production data: the clearest record of how they behave with users, software, and the real world. Judgment builds infrastructure for improving AI agents from production data.
English
Hari Subbaraj retweetledi
A small update - we had more traffic than anticipated. However, the endpoints are now scalable on Together AI for all models, including the 671B MoE.
Test out the model here: together.ai/models/cogito-…
(A huge thanks to the folks at @togethercompute for making this happen so quickly.)
Drishan Arora@drishanarora
Today, we are releasing 4 hybrid reasoning models of sizes 70B, 109B MoE, 405B, 671B MoE under open license. These are some of the strongest LLMs in the world, and serve as a proof of concept for a novel AI paradigm - iterative self-improvement (AI systems improving themselves). The largest 671B MoE model is amongst the strongest open models in the world. It matches/exceeds the performance of the latest DeepSeek v3 and DeepSeek R1 models both, and approaches closed frontier models like o3 and Claude 4 Opus.
English
Hari Subbaraj retweetledi
MISSING PERSON AT RISK; Katie Hong, 16 years old Asian female juvenile, approximately 5' tall and 100 lbs, brown eyes, has braces, and medium length dark brown hair. Last seen wearing a navy blue mid zip up jacket, grey sweats, and a red Patagonia backpack.
Katie Hong was last seen at her residence on Queens Ct at approximately 6:20pm on 07/25/2025.
Any information please contact Fremont Police Department 510-790-6800. In an emergency please call 911.
English
Hari Subbaraj retweetledi
chipmunk is up on arxiv!
across HunyuanVideo and Flux.1-dev, 5-25% of the intermediate activation values in attention and MLPs account for 70-90% of the change in activations across steps
caching + sparsity speeds up generation by only recomputing fast changing activations
English
Hari Subbaraj retweetledi
Our DeepCoder-14B LiveCodeBench v5 scores have been validated and put on the official leaderboard!
English
Hari Subbaraj retweetledi
🚀 Excited to partner with Meta to bring Llama 3.1 models to Together Inference and Fine-tuning! Up to 80 tokens per second for Llama 3.1 405B and up to 400 tokens per second for Llama 3.1 8B, which is 1.9x to 4.5x faster than vLLM while maintaining full accuracy.
together.ai/blog/meta-llam…
English
Hari Subbaraj retweetledi
I'm incredibly proud of the team at @togethercompute. Tons of deep research goes into this super fast inference engine: low-level GPU kernel optimization, smart quantization, and clever speculative decoding algorithms
Together AI@togethercompute
Today we are announcing a new inference stack, which provides decoding throughput 4x faster than open-source vLLM. We are also introducing new Together Turbo and Together Lite endpoints that enable performance, quality, and price flexibility so you do not have to compromise. together.ai/blog/together-…
English
@emma_jaeger I hop LI pays @KingJames well for this… got me to re-sign immediately
English
Hari Subbaraj retweetledi
One of my favorite things we've implemented @RunLLM recently: Automatically pulling in content from links in questions.
GIF
English
Hari Subbaraj retweetledi
We’ve been working on Aqueduct for over a year, and we’re super excited to share what we’ve been building: blog.aqueducthq.com/posts/aqueduct…
English
Hari Subbaraj retweetledi
Part 2 of @profjoeyg's reflections on his research career! Last week, he discussed the genesis of the ML revolution in powerful abstractions for building models. This post focuses on what comes after building models — using them in real applications: blog.aqueducthq.com/posts/the-real…
English
Hari Subbaraj retweetledi
Our co-founder @profjoeyg recently got tenure @Cal, and he took some time to write about how his career in ML research has been shaped by the broader trends in the space and the technologies that have enabled the machine learning revolution: blog.aqueducthq.com/posts/how-mach…
English
Hari Subbaraj retweetledi
1/ As some of you know, I recently got tenure at @UCBerkeley. I've been thinking about my research career and the simultaneous evolution of the ML space, and I wanted to share what I've been thinking about: @jegonzal/5732c3419c81" target="_blank" rel="nofollow noopener">medium.com/@jegonzal/5732…
English