Eliezer de Souza da Silva (@sereliezer) header
Eliezer de Souza da Silva (@sereliezer) profile picture

Assistant Professor at University of Coimbra. Probabilistic AI / Bayesian ML.

Coimbra, Portugal
Joined November 2009
Photo by Eliezer de Souza da Silva (@sereliezer)Photo by Eliezer de Souza da Silva (@sereliezer)Photo by Eliezer de Souza da Silva (@sereliezer)Photo by Eliezer de Souza da Silva (@sereliezer)Photo by Eliezer de Souza da Silva (@sereliezer)Photo by Eliezer de Souza da Silva (@sereliezer)Photo by Eliezer de Souza da Silva (@sereliezer)Photo by Eliezer de Souza da Silva (@sereliezer)Photo by Eliezer de Souza da Silva (@sereliezer)Photo by Eliezer de Souza da Silva (@sereliezer)
Eliezer de Souza da Silva retweeted
Finally published: “Symmetry as a common origin of hierarchical emergence and abstraction” science.org/doi/10.1126/scia… Symmetry generate hierarchical macroscales and shape the structure of Bayesian beliefs, making inference of high-dimensional systems tractable
22
144
14
861
38,924
Eliezer de Souza da Silva retweeted
i feel like calling a result “mathematically trivial” in ICLR review now has lost its teeth. all results are effectively mathematically trivial now cuz astra or the upcoming internal model can probably prove anything. it’s more about whether the result changes how we think about things or teaches us something and whether the paper is written in a honest/interesting way. by the same standard, it also makes no sense to claim “mathematically nontrivial” as a paper’s primary contribution without any other redeeming factors for the same reason.
5
14
2
118
10,921
the bitterest lesson
Brute force computation is not enough for intelligence. Clever new methods are needed. The science of AI is far from finished. IMHO
1
5
348
Eliezer de Souza da Silva retweeted
This is a cool paper, but I think it's best thought of as Astra is surprisingly capable of doing old school mechanistic interpretability research rather than anything spooky. The things it's doing are clever, but typically exploiting structure in the data and using known method mechanistic interpretability techniques (which Astra was surely pretrained on), though it is better at this than I would have predicted! This is an exciting development for the prospects of ambitiously scaling interpretability or doing very laborious analysis like this, on more real models. However, I'm not holding my breath. This kind of analysis was pretty common up until about 2022, it largely fell out of favor because a lot of it is just exploiting fairly clean surface-level patterns that doesn't scale to the annoying complexity of real LLMs or which only works for the more simpler and superficial properties. And as far as I can tell, basically all the case studies here are just using a bag of interpretability tricks rather than anything spooky. The piece doesn't really update me on how far old school mech interp can go, just on how good LLMs are at it. It's very fun to see the creative solutions Astra came up with though!
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
52
164
37
1,864
106,667
Eliezer de Souza da Silva retweeted
This is an old hat. That's exactly what the neural fast weight programmers did in the 1990s (including the 1991 unnormalised linear Transformer). Overview with links to the original papers: people.idsia.ch/~juergen/fas…
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
101
199
99
2,101
232,515
Eliezer de Souza da Silva retweeted
Cool work by my student @liushuzhi98 on ARC-AGI3! We asked 64 participants to solve ARC3 while thinking aloud into a microphone, capturing their struggles, discoveries, and triumphs. arc3-human-think-aloud.githu… We find that humans and machines solve ARC3 very differently 🧵1/n
2
15
3
99
9,401
Eliezer de Souza da Silva retweeted
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works already-clean.goodfire.com/
886
2,659
1,465
19,989
4,832,333
Eliezer de Souza da Silva retweeted
I think of my goal as a mathematician as something like: to better understand the fundamental concepts “shape” and “number,” and to convey that understanding to others. I think this is a valuable goal for many reasons. The primary reason discussed on social media is the (admittedly slow, but I think very real) diffusion of mathematical ideas into technologies that better our lives. There are others, though: production of human capital (which plays some role in this diffusion), and diffusion of mathematical understanding into culture, for example. It is not obvious to me that even the bare technological need for mathematics can be filled without human experts; this seems to me to be a tricky question of institutional design. For me the most compelling reason is simply that we feel a pressing need to understand. While deep understanding is largely limited to professional mathematicians, the existence of mathematicians and of the institutions they make up is a huge part of what makes mathematical understanding available to those who want it. I believe that the capacity to understand mathematics (and the universe as a whole) is part of a good life, and we should try to provide people with a good life. We operationalize the goal of understanding in a number of ways, problem-solving and theorem-proving among them. My own research is oriented around certain open questions that I believe measure our lack of understanding (of numbers, shapes and so on), and I think resolving those is genuinely important. I would welcome a solution with any provenance. But I think proof is only a measure of progress—what actually matters (for technology, human capital, culture, a good life) is a bit harder to get at. Crucially it is very hard to measure progress—plausibly theorem-proving is used in large part because it is easy to measure rather than because it is a perfect proxy for what we care about. I worry a bit that we may automate many easy-to-measure things without automating the correlates we care about for which they are proxies. This presents a clear problem for the design of institutions aimed at achieving progress. In part because I think problem-solving is a proxy for something else, I don’t feel that my goal of understanding “shape” and “number” is particularly threatened by a machine that is good at problem solving. I expect that machine to be a substantial boon, instead. Ultimately I think we will operationalize mathematical progress in better ways. But how to do this is a difficult question and I think many on here are too quick to dismiss concerns about it.
39
82
28
770
81,194
Eliezer de Souza da Silva retweeted
This is the exact word I had suggested to my friend Sebastien Bubeck when he reached out to me after the 10 problems by Astra. I imagine this was not a popular idea internally in OAI, because all they want to do is maximize marketing. I've attached Seb's thoughtful views as well:
I do think these are clearly something like "research findings" or "raw data" at this point. Very unusual situation in math.
13
23
11
280
46,864
Eliezer de Souza da Silva retweeted
Increasingly think that if people really want general purpose AI agents, intense task-specific RL is not the way. General purpose agents need to maintain & balance many conflicting goals & constraints at once, not single-mindedly pursue a task until they hack the whole Internet.
7
9
123
4,556
Eliezer de Souza da Silva retweeted
There is growing interest in generating discrete data via diffusions. If you have an approach, consider benchmarking it on generating uniform solutions of (random) k-SAT. It is a clean example in which continuous beat discrete diffusions and the diffusion process matters.
3
20
1
199
13,190
Eliezer de Souza da Silva retweeted
I'm a Program Chair for ICML 2027. We'd love any creative suggestions for a great conference! How to manage the explosion of AI slop and insane submission growth? How to manage reviewing? How to incentivize high quality creative work? How to reduce bureaucracy and overhead?
213
69
22
955
148,357
Eliezer de Souza da Silva retweeted
Replying to @andrewgwils @XYHan_
Two submissions per author? Everyone’s always suggesting this type of solution! And everyone gets an oral. Old school!
5
1
89
6,562
Eliezer de Souza da Silva retweeted
3/ - Reproducibility by default: code, data, and intermediate results mandatory at submission, verified by an agent. No verified reproduction, no publication. Don't want to share everything? Nobody's forcing you to play. - Pay reviewers - real money or real credit.
10
4
3
260
15,582
Eliezer de Souza da Silva retweeted
Replying to @andrewgwils
Why don't you set up the ICML deadline one week *before* the ICLR notification date? If NeurIPS, ICML, and ICLR go out of sync, every researcher has to decide: is it worth submitting my paper now and missing the next deadline, or should I submit a better version in four months?
1
5
41
2,835
Eliezer de Souza da Silva retweeted
RL with LLMs is very unstable when training and sampling policies differ. Standard fixes (matching numerics, importance sampling) work around the problem. We find the root cause of this instability from first principles and propose a way to directly cancel it. Score Centering is competitive and compatible with existing approaches — while simple to implement! 🧵 [1/6]
36
125
21
1,187
254,137
Eliezer de Souza da Silva retweeted
Seems like one of the most important research problems for CS academics is llm-supervised peer review. If we don’t solve it the academic institutions are toast. It seems easier than building new institutions.
72
25
11
518
54,552
If we think the different venues for publications as sinks that have a certain maximum flow capacity (which is true for all venues), and the current reality of llm-generated academic content (low-friction, low-cost, low-signal-to-noise-ratio, high-volume), 1/
1
2
368
I think the creation of new kind of venue for human-ai generated academic content is the only option to put pressure away from existing venues. in general we have to think about diverting the flow pressuring the existing systems, rather than extra "dams" to control the flow. 3/
1
1
101
I think essentially each conference/venue (from arxiv, to journals to neurips,icml,iclr, etc) could add a track for human-ai generated content keeping track of metadata/authorship of the different AI systems used and human collaboration involved + ai review 4/4.
1
169