award winning ai researcher (2 so far) | just graduated cs phd | author how to build conscious machines | musician | @bennettsrazor | i do not work @AnthropicAI
Brisbane, Australia
- Tweets 7,750
- Following 4,863
- Followers 16,090
- Likes 20,939
Pinned Tweet
happy to announce i just got examiner feedback on my thesis. recommended for acceptance as is, no changes, no corrections :)
155
322
3,399
252,568
my approach to life
ALT I don't remember who took this picture. Someone in Western Australia I think. If anyone remembers, tell me and I'll tag them.
3
1
30
631
Was feeling a bit demoralised this morning, and this tweet made me feel better.
One of the worst forms of brainrot AI has cultivated is pessimism about basic research, ie the idea that important work can only happen inside a frontier/neo lab and only with 10k+ GPUs, so the rest should not even bother. What a bleak way to think about science. And it's false.
2
2
60
3,487
people complain about AI music, but that's only because they haven't heard my neighbour's taste in music.
4
2
48
1,492
Michael Timothy Bennett retweeted
Freedom Causes Neural Generalisation
zenodo.org/records/22959791
✍️ @MiTiBennett via @ZENODO_ORG
👉 The author presents mathematical and empirical proof that selecting machine learning models for behavioural freedom maximises their probability of generalising to unseen data.
💡 By demonstrating that behavioural flexibility outperforms rigid constraints, this work reframes general intelligence as an inherent property of freedom rather than rule-following.
@Corix_JC @SonuMonika @TheAIObserverX @maponi @JagersbergKnut @ahier @ChuckDBrooks @Shi4Tech @CEO_AISOMA @sallyeaves @EchoofLight11 @rvp @dinisguarda @TarakRindani @FrRonconii @Ym78200 @Nicochan33 @Khulood_Almani @TysonLester @SusanHayes_ @EstelaMandela @JoannMoretti @sulefati7 @FernandaKellner @pchamard @NeiraOsci @darioandriani @ALouiseLavalett @Fabriziobustama @amalmerzouk @mikeflache @RLDI_Lamy @smaksked @Eli_Krumova @PVynckier @chidambara09 @AMontespan @sminaev2015 @NathaliaLeHen @MHcommunicate @Analytics_699 @DanielleLargier
ALT 📷 Source: @MiTiBennett via @ZENODO_ORG
4
19
28
934
wait wait, I can do better.
Here it is. This is the one: arxiv.org/pdf/2404.07227
cc: @MiTiBennett
Can an LM, starting from random init (!!), learn to generate all of its pretraining data?
Introducing Self-Play Pretraining with Zero Data. Two models start from random initialization: a generator proposes programs for a universal Turing machine and a learner trains on their outputs. We never train on any real data, but see predictable scaling on natural datasets: zero-shot val loss on images, text, audio, and melodies decreases predictably with self-play compute. And the learner develops in-context learning capabilities.
A fun proof-of-concept, co-led with @AdityaCowsik and @KfirDolev and co-authors @gbruno_dl, @ANourya @noahdgoodman, and @YoavLevine.
3
1
12
2,733
Michael Timothy Bennett retweeted
Probably one of the only researchers in AI I always read.
1
1
12
1,463
there is something very comforting about having a quite conversation when System of a Down is playing loudly
1
16
772
start your own colony of computational animals today.
thank you @MiTiBennett @drmichaellevin and @Plinz for inspiring me to build this
latticeanimal-production.up.…
7
10
60
2,481,529
AI slop is permission for me to write in an increasingly unhinged manner, to prove a human wrote it.
10
12
163
3,411
At the risk of grossly oversimplifying current frontier models, it makes a lot of sense that a lot of problems in math are low hanging fruit for AI. We’ve had proof systems for a long time, but now an LLM can translate words to math and evaluate the worth of a line of inquiry. It is in many ways a reflection of AlphaGo a decade ago, using search within rules, and model for the intuition that guides it. Likewise, a lot of what made up my non ML AI class was defining constraints and implementing A*, then putting a little learned heuristic into it. Now I guess we also use a model to learn the constraints (e.g. the math definitions). Yes, I know there is more to it (at least from my understanding of Deepseek V4, as I don’t know for certain what OAI has actually implemented internally), but this thread of where this came from makes sense to me.
We’re working with an independent advisory group of mathematicians to help OpenAI responsibly share advances in AI and mathematics.
The group will advise on how we assess and communicate new mathematical results, uphold academic and professional standards, and build tools that support mathematical research and learning.
Through this work, we want mathematicians to be at the center of shaping how AI supports mathematical understanding and how its benefits reach the wider community.
openai.com/index/advisory-gr…
8
7
90
6,476