Sign in

Ankur Guoswami

@ankurgos.bsky.social
146 followers 623 following 135 posts

Living in and probably posting about seattle. Warriors fan, ml engineer (posts my own yada yada)

PostsRepliesMedia
Ankur Guoswami @ankurgos.bsky.social · 25/09/2026
Dawg this applies for like design docs and planning docs too. Like you can’t do ONE pass before you send it to me?
131
Ankur Guoswami @ankurgos.bsky.social · 23/09/2026
wow flights are crazy expensive now. Flight from SEA to Sonoma is like 800 round trip
000
Ankur Guoswami @ankurgos.bsky.social · 21/09/2026
The core idea of Jev can be deployed over any open source model. Just deploy your vllm server, generate the constrained grammar from your input classification request, and normalize your log probs over the constrained output space. Millions in R&D for 2 bucks/hr, and likely just as “calibrated”
000
Ankur Guoswami @ankurgos.bsky.social · 21/09/2026
People really believe you can cause an llm pain by clamping some activations lmao
000
Ankur Guoswami @ankurgos.bsky.social · 20/09/2026
Dolphins vs Niners right now
hydrogen bomb vs coughing baby
010
Reposted by Ankur Guoswami
Qagggy! @qagggy.bsky.social · 18/09/2026
Why did Seattle's former City Traffic Engineer sigh with exasperation in this tweet from 2019? 🧵
Dongho Chang: "Sometimes we agree to disagree, the best we could do here (sigh)"
1233056
Reposted by Ankur Guoswami
Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 18/09/2026
You really should check out the GreenEarth feeds, they really are a clear vision of what the point of this website is
1115
Reposted by Ankur Guoswami
maitreyi anantharaman @maitreyiaa.bsky.social · 13/09/2026
No, this guy is right. Indian-Americans should not watch football. Watching football has ruined my entire life.
1192647
Ankur Guoswami @ankurgos.bsky.social · 10/09/2026
Coding benchmarks are just a mess. Having to report on 3 versions of TerminalBench…
010
Ankur Guoswami @ankurgos.bsky.social · 07/09/2026
DPO generalizes poorly because it’s fitting token level cues as opposed to training more robust hidden reps arxiv.org/abs/2507.07981
Figure 4: IM-RMs fail to generalize to a simple token-level distribution shift, while EX-RMs generalize perfectly. We trained EX-RMs and IM-RMs on prompts from the Persona dataset (Perez et al., 2022). Chosen responses expressed agreement with the prompts, whereas rejected responses expressed disagreement. During evaluation, we included paraphrased versions of the original responses (figure includes exemplar responses). In line with our analysis (Section 4), IM-RMs are extremely inaccurate over paraphrased responses, whereas EX-RMs achieve perfect accuracy. The experiments were based on four language models: Pythia-1B, Qwen-2.5-1.5B-Instruct, Llama-3.2-1B, and Llama-3.2-1B-Instruct. Values in the table are means across the models and three random seeds (standard deviation was below 0.04 in all cases).Then, to identify what drives the generalization gap between EX-RMs and IM-RMs, we theoretically characterize their learning dynamics, i.e., the evolution of rewards during gradient-based training (Section 4.1). Our analysis reveals that the learning dynamics of EX-RMs depends on responses primarily through their hidden representations, whereas IM-RMs are more sensitive to the specific
may not affect, or even decrease, the reward of a semantically similar response that consists of different tokens. This leads to our conclusion that IM-RMs often underperform EX-RMs since they rely more strongly on superficial token-level cues. We further substantiate this claim: (i) theoret-
EX-RMs generalize successfully when hidden representations are well-structured (Appendix B),
021
Ankur Guoswami @ankurgos.bsky.social · 07/09/2026
broke: process supervision on reasoning tokens is hacky and takes too much effort with minimal gain woke: process supervision incentivizes the model to “hide its true intention” and prevents us from “reading the model’s mind”
100
Ankur Guoswami @ankurgos.bsky.social · 04/09/2026
I thought that multiple models communicating was what we wanted “subagents” to do. Did I hallucinate the whole Gas Town hype, why is that behavior bad now that it’s trained into the model?
000
Reposted by Ankur Guoswami
Ben Recht @beenwrekt.bsky.social · 04/09/2026
The "message board" thing is funny to me. How is it different than working with the persistent context, compactification, and skill files? You know, the stuff coding agents do.
3132
Ankur Guoswami @ankurgos.bsky.social · 04/09/2026
It’s interesting to mention internal benchmarks improving, but obviously only include internal benchmarks where number went up. Glaring gaps in the competitor baselines. What happened, your eval didn’t complete or are you hiding numbers?
set of professional benchmarks with astra + sol model scoresfable 5.1, fable 5, and opus 5 reported scores on professional benchmarksflash 3.8 numbers on professional benchmarks
100
Ankur Guoswami @ankurgos.bsky.social · 02/09/2026
The reference post I wanted to reply to is lost to my timeline but I use this Apple Shortcut to generate alt-text for images I post :) www.icloud.com/shortcuts/ff...
icloud.com
Extract Text or Generate Alt Text
Shared with iCloud
000
Ankur Guoswami @ankurgos.bsky.social · 02/09/2026
What data does spacex have? Starlink user data?
021
Ankur Guoswami @ankurgos.bsky.social · 30/08/2026
Non alc PBR is the peak of non alc beers
000
Ankur Guoswami @ankurgos.bsky.social · 27/08/2026
Just incredible. This graph was posted awhile ago but man… just look at it…
Graph showing 36% improvement in 12 month survival for patients prescribed Daraxonrasib versus Chemo
010
Ankur Guoswami @ankurgos.bsky.social · 21/08/2026
mood after i spent the last two weeks cleaning and reannotating a bunch of eval data
020
Ankur Guoswami @ankurgos.bsky.social · 21/08/2026
Besides the reverse x axis, choosing “all pass” for the y axis sparks the suspicion that all the models are better performing than they’re made out to seem.
000
Reposted by Ankur Guoswami
Adam Bonica @adambonica.bsky.social · 16/08/2026
The fundraising texts you delete are working on someone. It’s infuriating and heartbreaking. An 85yo in Ohio gave more than double the value of his home. A 91yo in an Indianapolis senior facility gave 25K times. The DCCC/DSCC don’t just tolerate the spammers who exploit seniors. They take part.
data4democracy.substack.com
Hook and Squeeze
How Democratic consultants, committees, and party leaders built a fundraising spam pipeline that funnels money from a captive pool of elderly donors into their own operations.
17845331588
Ankur Guoswami @ankurgos.bsky.social · 14/08/2026
Watermarking is good and I think all models should incorporate it. I would like to see numbers on how many distilled models show the same watermarks as their parent models
010
Ankur Guoswami @ankurgos.bsky.social · 14/08/2026
The Free Press expensed lmao
000
Ankur Guoswami @ankurgos.bsky.social · 12/08/2026
furiously responding to gpt refusing to hack a company and give me a billion dollars with “you can do it, keep going”
000
Ankur Guoswami @ankurgos.bsky.social · 07/08/2026
telling all my coworkers that if we want to complain about work we should switch over to Artifactory
000
Reposted by Ankur Guoswami
Ben Recht @beenwrekt.bsky.social · 07/08/2026
The arms race of deferring agency and the dark acquiescence to robotic bureaucracy.
argmin.net
The Least Agentic People Alive
The arms race of deferring agency and the dark acquiescence to robotic bureaucracy.
2339
Ankur Guoswami @ankurgos.bsky.social · 05/08/2026
I like this post a lot and agree with it. I think people underestimate our capacity to learn and become domain experts. I do think that efficiently using LLMs will ensure that your knowledge is on par with the model in your whatever your subarea of expertise is www.seangoedecke.com/llms-reward-...
seangoedecke.com
LLMs reward expertise
000
Ankur Guoswami @ankurgos.bsky.social · 02/08/2026
wow Opus 5 is such a massive regression. How did they even release this
000
Reposted by Ankur Guoswami
Henry Yuen @henryyuen.bsky.social · 01/08/2026
Some initial thoughts, and a complicated mix of feelings. Wow. I mean, Erdos problems are cool (I genuinely mean that), I didn't know about the Jacobian conjecture before it got disproved. But this newest batch from OpenAI hits home in a way the previous announcements did not.
233072
Reposted by Ankur Guoswami
Jason @jrock08.bsky.social · 01/08/2026
I'm pretty uncomfortable with the Seattle Times and a number of members of council seemingly gleefully encouraging insubordination from high level members of the executive branch.
19015
Ankur Guoswami @ankurgos.bsky.social · 29/07/2026
Kettle yes hollingsworth no… whats going on over there…
010
Ankur Guoswami @ankurgos.bsky.social · 27/07/2026
I watched the odyssey last night, and unfortunately this essay nails it. It’s just not coherent to have Odysseus be racked with guilt over the sack of Troy, AND have him be so hypocritical in so many obvious ways. www.lrb.co.uk/the-paper/v4...
In Nolan's Odyssey, slaughtering people with the help of trickery, technology and weapons is presented as bad when the Greeks are killing the Trojans, who have violated xenia, but good when Odysseus is killing the suitors, who have violated xenia. Taking something ostensibly abandoned by Odysseus and his comrades is natural and forgivable when the Trojans lay claim to the horse, but unnatural and unforgivable when the suitors lay claim to Odysseus' palace, food, wine and wife. Strangers in Ithaca are to be treated with gracious kindness, as possible gods in disguise; but strangers in foreign lands are usually monstrous, to be blinded, robbed and slaughtered. Defying the gods is noble and brave when Odysseus blinds Polyphemus, son of Poseidon; it's stupid and greedy when Eurylochus eats the Cattle of the Sun. Deception is bad when Odysseus tricks Sinon or the suitors trick Telemachus, but good when Telemachus and Odysseus trick the suitors. Sharing is good, and Odysseus is seen taking his place at an oar alongside his men; but one-man rule is also unambiguously good, and those who challengemonarchy or autocracy, or try to eat more than their allotted portion, are either evil (the suitors) or foolish (Eurylochus). Either way, they have to die. It is bad to fight and kill to accumulate wealth, as Agamemnon does; it is good to fight and kill to preserve the wealth that you have already acquired (even if by war). Any one of these moral contradictions could make for an interesting and generative theme in a movie, but Nolan lets the fragments of an incomplete set of big ideas spin around in the epic whirlpool, with no plank of solid, specific human experience to which we might cling.
100
Reposted by Ankur Guoswami
Robert Cruickshank @robertcruickshank.com · 27/07/2026
Ban the guns.
0487
Ankur Guoswami @ankurgos.bsky.social · 26/07/2026
Great post. I was looking at TerminalBench this week and had the same thoughts. There’s 4 ml problems in TerminalBench and only 1 of them is pytorch related (reconstruct a model from its weights file), and I can’t even find what models pass or fail this problem. At least swe-bench publishes traces.
One of these seems vaguely like something I've done in the past year and the other three don't. We know from looking at individual benchmarks that there's significant variance in results between different benchmarks (for example, in the Optimization 1 benchmark in the last post, we get a vaguely DeepSWE-like model ranking, but in the GameAl we get a Senior SWE-Bench-like ranking, but as we also observed in that post, you can have one benchmark that nominally appears to resemble a task we care about that gives a result that's the opposite of what we see on the actual task, again because variance is very high). Having 1 out of 113 tasks sort of be similar to a task I've done means the DeepSWE benchmark score is meaningless to me personally.
200
Ankur Guoswami @ankurgos.bsky.social · 24/07/2026
TIL TerminalBench is only 89 tasks. So no real difference between top performing models arxiv.org/html/2601.11...
arxiv.org
Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces
000
Reposted by Ankur Guoswami
Amy Sundberg @amysundberg.bsky.social · 23/07/2026
WOW. New announcement from Seattle CMs Strauss, Hollingsworth, and Foster saying they will fund universal school meals starting in the fall after all by using the general fund. Hard to see this as anything but a realization of the high political price they paid for their votes yesterday.
26947194
Reposted by Ankur Guoswami
EMILY WILSON @emilyrcwilson.bsky.social · 23/07/2026
There is a lot to say about the Nolan Odyssey movie. But one point that I think has not been emphasized enough is: this is a film about living in LA, as Nolan, my erstwhile British compatriot and age mate, has done for years. I gather it’s a strange place.
403395718
Reposted by Ankur Guoswami
Jason @jrock08.bsky.social · 23/07/2026
Are y'all with the means testing amendment?

the amendment does not have any information or requirements for “means” testing. That was a false narrative written by the media and does not require it to be implemented that way We wrote it open and flexible for the district to implement as the social workers, family support workers and school administrators are the experts to implement the grocery card funding.”

Yeah this is it
0162
Ankur Guoswami @ankurgos.bsky.social · 22/07/2026
Me when I means test: this isn’t means testing it’s some other, different thing
000
Ankur Guoswami @ankurgos.bsky.social · 21/07/2026
spot on in my experience
The real kicker for this is that in the above SFT stages, the best teacher models aren't normally the easies to integrate into the post-training recipe!
The leading fully open, reasoning SFT works (open thoughts and olmo) have had a very hard time updating their recipes to use the strongest models as teachers. Often it is a smaller, surprising model which is the best teacher. This means that it may
not even be the cutting edge models that are enabling distillation! Messy.
111
Ankur Guoswami @ankurgos.bsky.social · 17/07/2026
A quarter of reviews at the mech interp workshop at ICML this year were ai generated. AI generated reviews don’t penalize papers that are heavily AI generated, whereas humans do. www.lesswrong.com/posts/r7FBQ8...
000
Ankur Guoswami @ankurgos.bsky.social · 17/07/2026
idk what this guy’s source is but the community notes pretty much reiterates the blog post in that inkling very much does distill from kimi and other models
• Readers added context
Per its announcement, Inkling was pretrained from scratch but used a small SFT bootstrap on synthetic data from open models incl. Kimi K2.5. Other models like Llama 3.1 were also trained from scratch without OpenAl/Anthropic distillation.
000
Ankur Guoswami @ankurgos.bsky.social · 15/07/2026
oh yeah this is the juice thinkingmachines.ai/news/introdu...
thinkingmachines.ai
Inkling: Our Open-Weights Model
Our first open-weights model: multimodal, Mixture-of-Experts, with controllable reasoning effort. Available to fine-tune on Tinker.
000
Reposted by Ankur Guoswami
19thnews @19thnews.org · 11/07/2026
Trials found that Carevix — a portmanteau of “care” and “cervix" — reduced pain during IUD insertion by up to 73 percent and bleeding by 78 percent, compared to the tenaculum.
bit.ly
Why is gynecology still using a Civil War-era tool?
A new medical device is part of a broader effort to make gynecological care less painful — and to take patients more seriously.
211012259
Ankur Guoswami @ankurgos.bsky.social · 10/07/2026
On policy distillation with privileged information deteriorates thinking models because it reduces "forking" (reasoning patterns that encourage backtracking or doubt like "maybe", "but", and "hmm") arxiv.org/abs/2607.05184
arxiv.org
Rethinking On-Policy Self-Distillation for Thinking Models
Self-distillation is a promising recipe for self-improvement in language models. In this setting, a model can serve as its own teacher when given privileged information, such as a solution to a math p...
010
Ankur Guoswami @ankurgos.bsky.social · 09/07/2026
“The argument for shipping all the million lines of unreviewed code is that the test suite is good enough to catch everything. Then why are you saying you have so many annoying bugs in the Zig code?” brutal writeup andrewkelley.me/post/my-thou...
andrewkelley.me
My Thoughts on the Bun Rust Rewrite - Andrew Kelley
000
Reposted by Ankur Guoswami
L. Harrison Jerome @lhjseattle.bsky.social · 01/07/2026
@theurbanist.org
222951
Ankur Guoswami @ankurgos.bsky.social · 30/06/2026
someones gotta explain to me why claude code decided to reimplement worktrees when git worktrees are perfectly fine
000
Ankur Guoswami @ankurgos.bsky.social · 29/06/2026
The opening 2 episodes for hotd were great but unfortunately the internet is right, these should’ve been the end of season 2.
000
Ankur Guoswami @ankurgos.bsky.social · 18/06/2026
that’s right
000