Pinned Tweet
New paper alert: "Latent Space Communication via K-V Cache Alignment" arxiv.org/abs/2601.06123
We propose a method for Large Language Models to communicate directly via their internal states, bypassing the need for discrete text generation 🧵
Lucio Dery Jnr Mwinm retweeted
I'm on the industry job market this year!
I'm a PhD student at CMU LTI. My work spans post-training, evals, alignment and multimodality!
Please reach out (DM) if you are hiring and think I'd be a good fit, I'd love to chat :)
Website: simran-khanuja.github.io/
CV: simran-khanuja.github.io/ass…
Lucio Dery Jnr Mwinm retweeted
On the job market this year! Open to faculty as well as RS/MTS positions and primarily interested in the science of language models. Please reach out if you are hiring and you think I would be a good fit.
Lucio Dery Jnr Mwinm retweeted
( P.S. I'm also on the industry job market, so please reach out if you find this work relevant :) )
Lucio Dery Jnr Mwinm retweeted
I’m very opinionated. So it’s nearly impossible for me to hear someone’s arguments and *not* find something I disagree with. (This includes reading things I myself wrote years - or weeks - ago.) And yet the amazing Grant Sanderson @3blue1brown has put out a long, detailed video on the future of mathematics, in which I cannot find a single flaw. I’ll be sending this to every PhD student and postdoc I know, and strongly recommend everyone watch it.
youtube.com/watch?v=LKrwhjOI…
(Of course I tried to say many of the same things in my conversation with Thomas Koberda. But Grant’s version, unsurprisingly, does an infinitely better job at communicating those ideas!)
Lucio Dery Jnr Mwinm retweeted
Everyone who says "why don't the AI companies work on things that are actually good for humanity?" - well, this is the sort of thing we *were* working on before ChatGPT came out and everyone dropped everything to compete with that instead.
Google Deepmind (FNV3) CRUSHED the forecast for #Isaias up to 11 days in advance. It picked up some appreciable signal on September 23, locked in on the exact landfall location4 days before it happened, and nailed the peak intensity consistently 3 days before it happened. A new age of forecasting is upon us.
First video: Track density using 1000 member ensemble
Second Video: Individual tracks and mean using 50 member ensemble
Lucio Dery Jnr Mwinm retweeted
When LLM demand exceeds capacity, which requests should go first?
Model providers already offer a few priority tiers. But even within a tier, requests compete for compute, and users value latency differently. Your coding agent might be blocked due to a request that belongs to an unattended background job.
Why not let the market decide?
In our new paper, Inference Auctions, users bid for request priority, and our auction allocates that priority to those who value it most.
Unfortunately, naively selling to the highest bidders destroys prefix-cache locality, increasing SGLang's average latency by up to 12x in our experiments.
Our fix is to run an auction over depth-first search traversals of the request prefix tree, ordering subtrees by average bid. This maximizes social welfare while retaining cache optimality, with VCG payments incentivizing truthful bidding, and the resulting request order preserves key performance metrics (cache hits and mean TTFT) of the underlying scheduler.
w/ @keegan_w_harris @sidpra5ad @nhaghtal and Michael Jordan
Paper: arxiv.org/abs/2609.40070
Project page: inferenceauctions.com
Lucio Dery Jnr Mwinm retweeted
Honoured to receive an outstanding paper award for our work! Couldn’t have done it without all my coauthors @KaiserWhoLearns @millicent_li @wordscompute @lltjuatja @JentseHuang and my advisor @gneubig 🫡
Congratulations to the authors of the 3 outstanding papers!
arxiv.org/abs/2607.01077
arxiv.org/abs/2604.08510
arxiv.org/abs/2604.13977
Lucio Dery Jnr Mwinm retweeted
I'm at #COLM2026 and on the academic job market this year! I'd love to chat about long context reasoning, distilling harness capabilities into architectural improvements, and better training methods for complex corpus tasks.
I'll be presenting work on Tuesday + Thursday ⬇️
Lucio Dery Jnr Mwinm retweeted
Second question on AI.
We are told repeatedly that AI is going to transform every aspect of our lives – jobs, productivity, inequality, science, communication, daily activities, social order, and politics, among others.
But this promise (or threat) is coupled with the rhetoric that such an important technology, with all of the risks and competitive pressures that it entails, should be left to experts or to “technocracy” (perhaps construed broadly to include some regulators).
These two statements are hard to reconcile in a democratic society.
If anything is half as important as AI is said to be (and I agree, AI is potentially very important and transformative), then involving democratic voice is essential. If something will shape our future in a democratic society, then its direction is for democratic institutions to decide.
My instinct is that democratic voice is essential, and relying too much on technocracy could be both dangerous and counterproductive.
The counterargument that AI’s direction can and should be entrusted to technocracy would go something along the following lines. First, democratic decision-making has become imperiled in our age of polarization. Second, AI is sufficiently complex that most citizens won’t have a deep enough understanding to meaningfully contribute to the debate (and even to the question of what we want from AI). Third, competition between different labs, and perhaps competition between the US and China, creates enough discipline for a socially beneficial direction of AI to be adopted. Fourth, today’s AI leaders are enlightened and ethical enough that within the framework created by competition, they can be broadly trusted.
There are many aspects of this counterargument that I do not find convincing. Taking them in order: polarization can be overcome, and big decisions and challenges sometimes bring societies together; in fact, delegating key decisions to technocracy without democratic input may diminish trust in institutions and experts, and may worsen polarization. Second, democratic voice does not require citizens to write code or design new models; the debate should be informative enough that citizens can weigh in about what type of future they want and how they trade off the costs and benefits of different options. Third, competition doesn’t seem to be a good disciplining framework; on the contrary, competition sometimes brings the worst out of both organizations and people. Fourth, if three decades of work on political economy and institutions has taught me anything, it is that we should not bank on the ethical grounding of unconstrained leaders.
But, still, I do not mean to immediately dismiss the technocracy option if there are more compelling arguments for it.
The question is, then, whether there are any circumstances under which such important decisions can be delegated to AI experts and technocracy.
One final secondary question: even if we managed to get democratic input in the United States or even in Europe, AI will shape the lives of everyone on this planet. How do we ensure that the voice of nearly 6 billion people who don’t live in the US, Europe and China also contributes to the debates on AI?
Nature research paper: Scalable decision-making for games of imperfect information
go.nature.com/3TvlQQS
If you're serious about AI accountability & looking to work on the institutional, legal and technical infrastructure it takes to make AI work for all of us - please consider applying to work with me at Stanford ❣️I'll be looking to recruit both phd students and postdocs this year
Lucio Dery Jnr Mwinm retweeted
Choosing an agentic memory architecture is a design problem that centers on how teams to structure context in their model harnesses.
We’re excited to release our MEMOIR benchmark to help teams evaluate how their approach to context management impacts performance on queries designed around longitudinal patient data.
We’re excited to share MEMOIR, an open-source benchmark for evaluating representations of longitudinal clinical memory.
MEMOIR: Memory Evaluation over Multi-year Oncology Inference & Retrieval includes 117 synthetic oncology patients, 65,377 clinical events, and 3,617 questions spanning factual retrieval, rare-event detection, clinical episode reconstruction, and longitudinal reasoning.
Clinical memory has a different structure from conversational memory. A patient’s history is assembled from events across many systems, at very different frequencies, over months or years. Answering a question may require finding a single event, reconstructing a time series, or connecting observations across an entire treatment episode.
We evaluated several memory and retrieval approaches, including vector-based memory (mem0), temporal knowledge graphs (Graphiti), agent-managed memory (Letta), direct retrieval over Olira’s typed clinical events, and materialized patient state (Olira State).
The clearest differences emerged when an agent had to reason across time. Reconstructing a time series, connecting events across a treatment episode, or understanding how a patient changed over time made the underlying representation much more consequential.
Blog: olira.ai/blog/the-wrong-shap…
GitHub: github.com/olira-ai/memoir-b…
Lucio Dery Jnr Mwinm retweeted
I must be among an extremely small group of people (n=1?) that have both 1) trained a frontier LLM and 2) designed and synthesized custom viruses in a lab with my own two hands.
And I think that the takes on AI killing us all by creating dangerous viruses is total bogus.
here's the argument some ai lab employees are making
- they genuinely believe what they're working on might kill everyone
- they continue to work on it because if this thing gets built it should be by responsible people like them
- the reason they do money things like selling enterprise deals and IPOing is solely to fund this effort
if this is all true then why is employee equity a thing? if you want to tell this story of being terrified of what you're building it doesn't work when there's a clear profit motive
it's particularly weird when people vest and then leave and write a tirade about how these companies are bringing the end of everything
if you want to act like this is the manhattan project, those scientists didn't have equity in something that gives them generational wealth
you don't get to say something bad is happening while profiting off it and also being seen as the good guy. one of these has to go
Lucio Dery Jnr Mwinm retweeted
Replying to @AgustinLebron3
Not really observations that others before me haven’t made, but:
- Despite the seemingly magical nature of LLMs, reflection over a >3 month timescale suggests my total productivity hasn’t increased by over 100%, or perhaps even by over 50%, and a lot of time is actually wasted because LLMs enable me to spend time on gratifying but low-productivity tasks that in the future turn out to not be useful
- Also, capabilities are incredibly spiky and highly correlated with the degree of investment poured into them, which my earlier tweet about math benchmarks implicitly points out
- From the above, it seems that the nature of LLM intelligence is wildly dissimilar to that of human intelligence and we won’t trivially get to something superior to human intelligence in all important respects just by scaling up existing approaches with various tweaks; even if AGI Is eventually achievable, this implies a significantly longer timeline
- Benchmark progress is almost definitionally guaranteed to happen because the process of constructing a benchmark is a direct precursor to the process of constructing a training dataset used for hill climbing that benchmark, but the scope of what can be captured in a benchmark is (at least for now) grossly lacking in terms of its relevance to real-world work, with maybe several limited exceptions
- Progress seems highly gated by data but the nature of model training means that each “next dataset” is significantly harder to assemble than what preceded it; some wins are possible through synthetic methods but those feel more like “patching up gaps” than “pushing the frontier forward”
At a higher level, I guess I’d say there’s a sort of refusal to think carefully about what models are or are not useful for in a rigorous way which I find personally quite annoying, and instead a reliance on some nebulous notion of being “AGI pilled” as a replacement for serious thought. I think people are very quick to anthropomorphize LLM intelligence because humans communicate through words and we infer the intelligence of human counterparties through comprehension of their language, but this leads them to wrong conclusions; for example if we observe that a new model proved some incredible mathematical theorem, some will say, “well, don’t we have AGI now, huh?” But to me, it’s actually more like, “well, given how hard it would have been for a human to do these mathematics, and given the limited economic effect of LLMs upon the world so far, isn’t it actually a negative datapoint vis-a-vis the generality of LLM intelligence?”
Lucio Dery Jnr Mwinm retweeted
I'm reflecting on how much research has changed since I've joined the PhD and wrote a short blog post about it (I joined in the tail end of the BERT era!). It seems pretty crazy how different processes are now, and I took the chance to do a retrospective before graduation:
nightingal3.github.io/blog/2…
Checkout our latest work on instruction steered parameter synthesis. This work is a step towards the North Star of achieving flexible and accurate inference time generation of model weights that target specific tasks / capabilities.
0/ With @derylucio, we co-led the development ofcSkillSmith, an LLM that can perform instruction steered parameter synthesis and composition by treating weights as a native modality. SkillSmith outputs PEFT parameters for a target task given text knowledge and related task weights 🧵
Lucio Dery Jnr Mwinm retweeted
I'm so happy to be able to announce the first general purpose sign-language-to-text translation (SL2T) model from my team that's out today and powering a new ASL input feature on @Android phones (and don't worry, we're not stopping at ASL).
[CL] SkillSmith: Learning to Compose Parametric Skills and Textual Knowledge
L M. Dery, B A Tjandra, S Samiei, A Kuncoro… [Google DeepMind] (2026)
arxiv.org/abs/2607.27497





































