@sgdescent

Interested in making LLMs go brrr x+N: @datologyai x: @LTIatCMU x-N: @ProseMsft @Zomato @OpenAI

 Pittsburgh
Joined February 2021
  • Tweets4.2K
  • Following1.8K
  • Followers1.3K
194Media
Feeling insanely burnt out, but when I open twitter it makes me feel bad about trying to rest while everyone races towards the next big thing. What a weird feeling
8
341
Why is this lesson so hard to internalize tho? You are going to make amazing things at a faster pace, and solve problems at a higher level of abstraction than before. This has always been the case with most innovations in tech, just the scale is now different!
Scott
@scottsttsSep 25
My god this is such a good speech that every SWE needs to hear. You know what? Every person should hear it Keep the happy memories, eyes on the reality, be excited about the future. That’s the best that anyone can do
1
2
332
Saksham retweeted
JosH200
@josh_willsSep 22
I physically cannot bring myself to use an Opus model again the ick from 5 is just that strong
36
6
3
260
23,109
Can someone explain to me how Jev works?
159
DM US
Lawrence Jang
@JangLawrenceKSep 11
We are running a weekly Agents + RL + Envs seminar at CMU this semester. We are happy to field external speakers, please DM me if you are interested in giving a talk! We are also looking for sponsors - please DM or email me. You can also reach out to @kohjingyu @PranjalAggarw16 @stephofx @ianwu97!
2
485
I’m seeing all these cool Astra robot-arm demos after @cdngdev’s tweet, but I’m curious about the interface: Is Astra writing/executing scripts to control the robot, or issuing discrete low-level commands step by step?
254
Saksham retweeted
Ari Morcos
@arimorcosAug 31
More and more people are waking up to the reality that frontier level custom models are going to become the standard in the coming years. Thomson-1 is just the first example. And note, they didn't spend $40M on the model, but rather on the model factory. The model itself only cost $450k. Now, it can be updated as new open source models come out to always keep Thomson Reuters at the frontier. And it was built on data curated by @datologyai.
Zach Moskow
@zachmoskowAug 31
CRAZIEST AI story of the week. Thomson Reuters just launched an LLM called Thomson. It starts with Qwen3.5-397B, then is continually trained on 175 years of proprietary legal, tax, accounting, and news data. They spent $40M on it. Two years ago, Thomson Reuters bought a pre-revenue legal-AI startup called Safe Sign. That team built Thomson. On Thomson Reuters’ published benchmarks, they say it is comparable to Claude Opus 4.8 and ahead of GPT-5.5, Claude Sonnet 5, and Gemini 3.1 Pro overall. They have used less than 10% of their data collection so far. We’re going to see this a lot more often.
8
20
4
206
42,823
Hmmm, something weird happened @ChatGPT; I had Pro-Gratis 20X Pro, which was supposed to end on 9th September. Using the student promo downgraded me. Seems like an edge case? CC: @thsottiaux
1
362
My friend (sales expert, business owner) and myself started a small local AI consulting company. Not sure if we will make any real money but excited about this
2
1
48
1,478
I'd recommend touching on following areas to get a better sense of how production inference works - Roofline analysis to determine compute, memory-bound - Changes in model architectures and motivation e.g. DeepSeek Sparse Attention - When manual kernels are needed vs. when to use vendor library kernels such as FlashInfer - How kernel fusion works and common fusion patterns - Until which point ML compiler optimizes and when manual optimization takes over - Data, tensor, pipeline parallelism and collective communication overhead - Characteristics of agentic coding workload and which optimizations to focus on - CUDA graph, PDL, etc.
Archie Sengupta
@archiexzzzAug 17
learn about inference engineering as fast and deep as you can
6
12
206
10,205
Spent hours digging into GEPA logs finding ways to fix your run? 👀 We have reimagined GEPA to be human-steerable. We built deep observability tools allowing agents to monitor, proactively surface, and fix issues in your run. We provide multiple levers for users to intervene and meaningfully affect the optimization trajectory! Early access signup in 🧵👇🏽
Previewing a screenshot of the interface with a four pane view. The leftmost pane is a candidate explorer showing the lineage of evolved candidates. The center shows an overview of the run, placing candidates on an annotated score over metric calls chart. The bottom pane shows GEPA Agent in action making a proposed fix to improve optimization. The rightmost pane shows the best candidate.

ALT Previewing a screenshot of the interface with a four pane view. The leftmost pane is a candidate explorer showing the lineage of evolved candidates. The center shows an overview of the run, placing candidates on an annotated score over metric calls chart. The bottom pane shows GEPA Agent in action making a proposed fix to improve optimization. The rightmost pane shows the best candidate.

2
13
2
82
19,050
Saksham retweeted
Vansh
@threeoneonetwoAug 11
Indian tech bros treat going to SF as if they’re doing hajj 😭😭😭
99
323
59
7,681
291,935
Saksham retweeted
Akshat Bubna
@akshat_bJul 29
We’re hiring ML research interns! If you’re a current PhD student interested in problems like adaptive speculation, off-policy RL, elastic rollout fleets and more, come talk to us: jobs.ashbyhq.com/modal/38888…

ML Research Intern

Research Intern

jobs.ashbyhq.com
11
27
2
544
61,135
New blog post: NVFP4 GEMM on Blackwell. Implemented NVFP4 GEMM for B200 in CUDA and wrote about many of the well known techniques used in Blackwell GEMMs, along with CLC, PDL, and selecting parameters. Code available as well! aramesh10.github.io/nvfp4-ge…
1
2
13
805
Recently, I've been helping organize a small ML Systems reading group @CMU. My first presentation was on performance principles for LLM performance, and I've decided to release it publicly! Hope it's helpful! sidbabu.com/posts/llm-perf-m…
2
3
1
9
758
Saksham retweeted
Amy Tam
@amytam01Feb 17

The Cost of Staying

Article · @amytam01
191
912
283
8,255
2,869,688
It is the nature of products with ML baked in where you see such failures. Even if new grads don’t have business sense (which they do?) there are other systems in place that make sure no product goes out with this intention, it happens and they will fix it. NL search is hard
Ankit Jxa
@kingofknowwhereMay 11
This is what happens when you hire new grads with zero business understanding to do ML products. if cosine similarity >= 0.85 show result (doesn't do Entity vs misspellings vs brand name check) 😭
1
3
763
Spending billions to train the "best" base model? You might be optimizing the wrong thing! 🎯 We show that controlling sharpness during mid-training leads to over 35% less forgetting after fine-tuning / quantization... even when the base model itself gets worse. 🧵 Takeaways for pretraining: - Use SAM (Sharpness-Aware-Minimization) in the final steps (~10%) - Try much higher learning rates (yes, even ~10× larger) 1/9
29
90
17
615
599,710
Saksham retweeted
Bhutani
@justbhutaniApr 10
We are hiring in SF in robotics. Folks who wanna live and work end to end - we call you to reach out to us.
Kalshi Finance
@Kalshi_FinanceApr 9
Stanford CS graduating class of 2026 just got their final placement statistics Out of 312 graduates: 18 have full-time offers That's a 5.8% placement rate from the most prestigious CS program in the fucking world 2019 placement rate was 94%. 2022 was 78%. 2024 was 31%. Now this. The other 294 are fighting over 47 internships that require "3+ years production experience" Career services is telling them to "consider adjacent fields" while the department just took a $50M donation from a company that replaced 2,400 engineers with Claude One kid showed me his rejection tracker: 1,247 applications since September. 12 phone screens. Zero offers. His parents refinanced their house for his tuition The career fair had 8 companies and 300 desperate students in $180k of debt Meanwhile the CS department just announced they're expanding their PhD program because "industry demand for AI research has never been higher" The same week they sent acceptance letters to 89 new undergrads These kids thought they were learning to be engineers. Turns out they were training to be obsolete.
3
7
1,897
One of my really goated friends @SCSatCMU is looking for an ML Research/Engineer intern for the summer rajworking.github.io/. @Dogged_Raj His offer got rescinded due to unfortunate circumstances. Hire him!
2
2
33
4,720
Load more