xentloss (@xentloss) profile picture

underparameterized nlp enjoyer

Joined June 2022
Photo by xentloss (@xentloss)Photo by xentloss (@xentloss)Photo by xentloss (@xentloss)Photo by xentloss (@xentloss)Photo by xentloss (@xentloss)Photo by xentloss (@xentloss)Photo by xentloss (@xentloss)Photo by xentloss (@xentloss)Photo by xentloss (@xentloss)
Damn that’s crazy. Boner fans use code ROCK-HARD-LOCK-IN for $10 off your first order.
Breaking news… my boners are now longer than 4 hrs Longer than > Titanic > Lord of the Rings (extended edition) Stats: > 4 hr 2 min > 5 episodes > 93% strength Out of database of 55,000 measurements: > 99th percentile among all males > 93rd percentile among men 20-25 years of age > 100% percentile among men my age > only 230 sessions exceed 4 hrs Nighttime erections (NTE) are a tier 1 longevity biomarker as critical to systemic health as VO2 max, resting heart rate, HRV, and blood pressure. Women have them nightly too. NTEs decline with age. Mine have been getting stronger over the past 18 months. This is a personal best. Low night time erection scores mean the following: > 2x risk of heart attack and stroke in next 4 years > early warning for cardiovascular problems > poor nervous system regulation Protocol for strong NTEs: > high sleep quality > cardiovascular fitness > tadalafil 5 mg - I take this longevity medication daily > calm nervous system
2
2
1,130
fascinating latex template optimized for mobile scrolling
We are releasing our speculative decoding approach, together with the draft model for Qwen3.5-27B and the reference implementation in SGLang. Paper: arxiv.org/abs/2607.06763 Model: huggingface.co/trymirai/weav… SGLang implementation: github.com/trymirai/sglang
1
25
4,039
Cool Search paradigm by @jobergum: “BM25 + Grep is all you need”: 1. Use BM25 to narrow down a large universe of documents to a few candidates 2. Expose the candidates to the agent in a virtual file system 3. Let the agent grep through them and find information like it normally would on a computer
2
13
4
107
35,847
off policy evaluation, active testing, hoeffing races
78
workflows clicked for me as durable coroutines. PTC programs have the same shape, except they're authored by the agent at runtime. this begged the natural question — why shouldn't they execute durably too?
Blog: The Agent Is a Workflow That Writes Itself How we productionized the RLM with durable execution. Subagents lower to child workflows and PTC runs through a deterministic workflow-space interpreter. Every tool, subagent, and PTC call goes through a single recursive dispatch loop. Closure under replay, retry, cancel. We call these durable agents.
1
1
7
545
much of this is downstream of @a1zhang and @lateinteraction's RLM work. all harnesses are headed in this direction, and the composition with durable execution is just beautiful.
1
131
pomdp fuck yeah
The Context Gathering Decision Process: A POMDP Framework for Agentic Search @ChinmayaKausik et al. at Netflix formalize iterative agentic search as a POMDP and introduce a predicate-based belief state, improving multi-hop reasoning and saving tokens. 📝 arxiv.org/abs/2605.07042
133
really like this task
We set out to build a better retriever, so we looked for the hardest IR benchmarks. For each, we asked how much headroom remained by running oracle reranking with a frontier LLM. Most had little room left! So we built OBLIQ-Bench to study much harder search queries than before.
1
8
3,680
xentloss retweeted
crazy decision to line the founders up like this
We raised $20M from @sequoia to redefine how software is deployed.
44
9
4
1,249
127,940
how old were you when you thought about swiss cheese as a joint density over arrival time and compact sets then how old were you when you realized stacking pieces of cheese is just poisson superposition
1
2
8
271
bidirectional encoders with extra steps
LLMs process text from left to right — each token can only look back at what came before it, never forward. This means that when you write a long prompt with context at the beginning and a question at the end, the model answers the question having "seen" the context, but the context tokens were generated without any awareness of what question was coming. This asymmetry is a basic structural property of how these models work. The paper asks what happens if you just send the prompt twice in a row, so that every part of the input gets a second pass where it can attend to every other part. The answer is that accuracy goes up across seven different benchmarks and seven different models (from the Gemini, ChatGPT, Claude, and DeepSeek series of LLMs), with no increase in the length of the model's output and no meaningful increase in response time — because processing the input is done in parallel by the hardware anyway. There are no new losses to compute, no finetuning, no clever prompt engineering beyond the repetition itself. The gap between this technique and doing nothing is sometimes small, sometimes large (one model went from 21% to 97% on a task involving finding a name in a list). If you are thinking about how to get better results from these models without paying for longer outputs or slower responses, that's a fairly concrete and low-effort finding. Read with AI tutor: chapterpal.com/s/1b15378b/pr… Get the PDF: arxiv.org/pdf/2512.14982
1
220
let’s just take peptides and do TRT and get jacked and raise a series B
5
159
a hospital has 2 cancer wards, A and B. due to recent policy changes, the hospital is forced to transfer some patients from ward A to ward B. conveniently, they can derive from an oracle that this move will lower the mortality rate of both wards (1/2)
Replying to @VahidK @xai @OpenAI
let’s just say the avg iq of both companies went up
1
4
738
but, did this policy change actually save lives? what kind of oracle would you need to infer that it did or didn't? and what could you conclude in the absence of such an oracle? (2/2)
1
2
202
these guys are so fun
Introducing Digital Red Queen (DRQ): Adversarial Program Evolution in Core War with LLMs Blog: sakana.ai/drq Core War is a programming game where self-replicating assembly programs, called warriors, compete for control of a virtual machine. In this dynamic environment, where there is no distinction between code and data, warriors must crash opponents while defending themselves to survive. In this work, we explore how LLMs can drive open-ended adversarial evolution of these programs within Core War. Our approach is inspired by the Red Queen Hypothesis from evolutionary biology: the principle that species must continually adapt and evolve simply to survive against ever-changing competitors. We found that running our DRQ algorithm for longer durations produces warriors that become more generally robust. Most notably, we observed an emergent pressure towards convergent evolution. Independent runs, starting from completely different initial conditions, evolved toward similar general-purpose behaviors—mirroring how distinct species in nature often evolve similar traits to solve the same problems. Simulating these adversarial dynamics in an isolated sandbox offers a glimpse into the future, where deployed LLM systems might eventually compete against one another for computational or physical resources in the real world. This project is a collaboration between MIT and Sakana AI led by @akarshkumar0101 Full Paper (Website): pub.sakana.ai/drq/ Full Paper (arxiv): arxiv.org/abs/2601.03335 Code: github.com/SakanaAI/drq/
1
225
from >700ms to this w opus 4.5
2
156
still finding excuses to write proofs after graduating
2
152
:.2f
playing around with cudagraphs: they do bring nice speedups if we keep their limitations aside for a moment
2
207
7) Now come probabilistic computers. The idea itself was not invented by Extropic. See for example arxiv.org/abs/2108.09836 In short, you can build an analog computer which does the Gibbs sampling directly in hardware. This allows you to accumulate samples very efficiently
1
1
1
25
1,883