banner
inference.vc
Ferenc Huszár
@inference.vc
2.1K followers 330 following 200 posts
Secular Bayesian. Professor of Machine Learning at Cambridge Computer Lab Talent aficionado at http://airetreat.org Alum of Twitter, Magic Pony and Balderton Capital
airetreat.org
Posts Media Videos Starter Packs
inference.vc
Ferenc Huszár @inference.vc · 1d
I’m curious what should/could they have done which in hindsight would have worked?
1
inference.vc
Ferenc Huszár @inference.vc · 1d
This seems to assume NGOs in those countries actually developed any productive counterstrategies or actionable insights. My impression in Hungary is that they pretty much barely hold on. I’m not sure what there is to learn other than getting a preview of what’s coming.
1
inference.vc
Ferenc Huszár @inference.vc · 6d
2 4
inference.vc
Ferenc Huszár @inference.vc · 6d
@nsaphra.bsky.social thank you f or your service, keep up the good work
2 1
inference.vc
Ferenc Huszár @inference.vc · 11d
I like the sound of this common EU corporate entity idea!
Except quite clearly the virtual state shouldn’t be called “regime 28” but “regime 0” or failing that, “regime -1”.

Open for consultation until 30 Sept so take a look!

www.eu-inc.org
EU–INC — One Europe. One Standard. — Pan-European legal entity.
EU–INC is a proposal for a pan-European standardized legal entity to unlock pan-European startup scaling.
www.eu-inc.org
1 4
inference.vc
Ferenc Huszár @inference.vc · 12d
Like a phoenix rising from the ashes,
like a turd that wouldn’t flush,
I re-emerge again in the geriatric Millennial basketball scene.
Wish me good health.
5
inference.vc
Ferenc Huszár @inference.vc · 20d
Mom, can we watch Star Wars?
Mom: We have Star Wars at home

Star wars at home:
inference.vc
Ferenc Huszár @inference.vc · 28d
We will have GPT6 before GTA6
2 13
Reposted by Ferenc Huszár
teorth.bsky.social
Terence Tao @teorth.bsky.social · Aug 18
I wrote an op-ed on the world-class STEM research ecosystem in the United States, and how this ecosystem is now under attack on multiple fronts by the current administration: newsletter.ofthebrave.org/p/im-an-awar...
I’m an award-winning mathematician. Trump just cut my funding.
The “Mozart of Math” tried to stay out of politics. Then it came for his research.
newsletter.ofthebrave.org
19 310 750
inference.vc
Ferenc Huszár @inference.vc · Aug 8
OpenAI automated HR assistant
4
inference.vc
Ferenc Huszár @inference.vc · Jul 21
Generalised second price auction for JavaScript imports?
2
inference.vc
Ferenc Huszár @inference.vc · Jul 21
Arrived in Sarajevo for @eemlcommunity.bsky.social
Located in a narrow valley, it’s an ideal setting to get the momentum going in my deep learning talk later.

I’m keen to explore beyond the hotel coffee though which, one assumes, is carefully curated to highlight floral notes of a morning cigarette
6
Reposted by Ferenc Huszár
ab-carrell.bsky.social
Annabelle Michael Carrell @ab-carrell.bsky.social · Jul 19
So you want to skip our thinning proofs—but you’d still like our out-of-the-box attention speedups? I’ll be presenting the Thinformer at two ICML workshop posters tomorrow!

Catch me at Es-FoMo (1-2:30, East hall A) and at LCFM (10:45-11:30 & 3:30-4:30, West 202-204)
ab-carrell.bsky.social
Annabelle Michael Carrell @ab-carrell.bsky.social · Jul 14
Your data is low-rank, so stop wasting compute! In our new paper on low-rank thinning, we share one weird trick to speed up Transformer inference, SGD training, and hypothesis testing at scale. Come by ICML poster W-1012 Tuesday at 4:30!
lestermackey.bsky.social
Lester Mackey @lestermackey.bsky.social · Feb 18
New guarantees for approximating attention, accelerating SGD, and testing sample quality in near-linear time
4 5
inference.vc
Ferenc Huszár @inference.vc · Jul 16
Evidently.
1
inference.vc
Ferenc Huszár @inference.vc · Jul 16
BREAKING: OpenAI point guard Jason Wei joins Meta in shocking trade news. OpenAI receives two ConvNet researchers and a 2027 first-round draft pick. Veteran Yann LeCun rumored to be missing another season due to slow Achilles recovery. Apple “cautiously optimistic” about G-league chances.
2 1 41
Reposted by Ferenc Huszár
ab-carrell.bsky.social
Annabelle Michael Carrell @ab-carrell.bsky.social · Jul 14
Your data is low-rank, so stop wasting compute! In our new paper on low-rank thinning, we share one weird trick to speed up Transformer inference, SGD training, and hypothesis testing at scale. Come by ICML poster W-1012 Tuesday at 4:30!
lestermackey.bsky.social
Lester Mackey @lestermackey.bsky.social · Feb 18
New guarantees for approximating attention, accelerating SGD, and testing sample quality in near-linear time
https://arxiv.org/abs/2502.12063
2 7 25
inference.vc
Ferenc Huszár @inference.vc · Jul 14
Low-Rank Thinning is presented tomorrow at ICML

⏰ Tue 15 Jul 4:30 - 7 p.m. PDT

icml.cc/virtual/2025...
2
inference.vc
Ferenc Huszár @inference.vc · Jul 12
Off to ICML next week?

Check out my student Annabelle’s paper in collaboration with @lestermackey.bsky.social and colleagues on low-rank thinning!

New theory, dataset compression, efficient attention and more:

arxiv.org/abs/2502.12063
Low-Rank Thinning
The goal in thinning is to summarize a dataset using a small set of representative points. Remarkably, sub-Gaussian thinning algorithms like Kernel Halving and Compress can match the quality of unifor...
arxiv.org
5 11
inference.vc
Ferenc Huszár @inference.vc · Jul 11
They solved the alignment problem
techmeme.com
Techmeme @techmeme.com · Jul 11
Further tests reveal Grok 4 appears to consult Elon Musk's views on a range of sensitive topics, and its answers tend to align with Musk's personal opinions (Maxwell Zeff/TechCrunch)

Main Link | Techmeme Permalink
1 1 14
inference.vc
Ferenc Huszár @inference.vc · Jul 10
Ah, what I meant is restricting inputs to A!=B, so say I don’t care about the diff between > and >=. But in any case I will read the symmetry argument properly.
1
inference.vc
Ferenc Huszár @inference.vc · Jul 9
Career highlight for sure.
1
inference.vc
Ferenc Huszár @inference.vc · Jul 9
Yes, and he didn't mind. I read it somewhere that he doesn't like men except Obama, so I felt pretty proud.
1 1
inference.vc
Ferenc Huszár @inference.vc · Jul 9
There once was a good man called Frederik
He landed a sweet job in Limerick
For fifty-six grand
He would teach on command
"Entschuldigung", "Tschüss" and "Ich liebe dich"
10
inference.vc
Ferenc Huszár @inference.vc · Jul 9
I don't think we disagree. You're saying there is no guarantee, and that it is not hard to come up with examples demonstrating it fails. I say there's no guarantee, which is why it's very surprising and intriguing that there are non-trivial examples where it does not fail.
1
inference.vc
Ferenc Huszár @inference.vc · Jul 9
I generally agree, of course. But your example is 'no-free-lunchy' in that you really care about the A=B cases and no general-purpose prior or inductive bias can fill that gap if you've only seen A!=B demonstrations. If the inputs are always A!=B, learning 1[A>B] exactly could be sample efficient.
2