#BitterLesson - Twitter Hashtag

20 days ago

or these systems will implicitly emerge inside a single model with enough scale and proper data engineering #bitterlesson

Dmytro Hrybov

@dimentary

20 days ago

right now, systems 1 and 2 in robotics communicate with each other mainly through text, one example is task planners prompting VLAs to execute subtasks i think this will be a major bottleneck, it’s hard to express richness of the physical intent with text, if you think about it e.g. when we try to grab a cup, our exact actions depend on whether it was hot/slippery/damaged somewhere/many other things the way our brains “handoff” this information from system 2 to system 1 during actions feels way more integrated

4

38

1

25

5K

2

9

0

1

1K

Kostas Daniilidis

@KostasPenn

4 months ago

@GeorgiaChal @jon_barron @vincesitzmann @anand_bhattad #bitterlesson The low-dimension of the action space does not make the problem easy. What matters is the structure of the latent manifold and what this encodes. 3D representations are efficient because they are invariant to nuisances like viewpoint, illumination, and appearance in pixel spaces. With infinite pixel2action episodes from behavior cloning or RL these representations may eventually emerge the way they emerged during evolution (probe animal brains and you will see equivariant 3D representations, work by Logothetis, Poggio and others). If the task were as simple as @jon_barron claims, no 3D or flow or stereo would have emerged in biological brains. Until we operate in evolution-like gpt-like data, hard-coding geometric structure is the best inductive bias we can apply. Evolution converged to geometry.

Georgia Chalvatzaki

@GeorgiaChal

4 months ago

@jon_barron You’ve put your finger on the real question there… where structure enters, modular vs end-to-end... That’s exactly what we’re debating, so I’ll save the long response for the panel. 🙂

0

5

0

1

9K

3

64

15

29

14K

Kechen Li@LinChance

@Kechen_Li2018

9 months ago

Pure Bitter Lesson vibes: skip hand-crafted experts, let search + verifiable feedback + compute win. Coding agents aren't just tools—they might be the fastest path to true general agents. What do you think? 🤖 #AI #AGI #CodingAgents #BitterLesson

Kechen Li@LinChance

@Kechen_Li2018

9 months ago

The killer feature? Scalability and extensibility. Persistent memory + unlimited execution cycles = an "infinite external brain." Suddenly time, complexity, tools, even self-improvement all scale in ways pure LLMs never could. 🚀

1

3

0

298

0

5

0

202

Joscha Bach

@Plinz

about 1 year ago

There is a ton of fascinating and inspiring work by brilliant researchers at CogSci 2025, but the bulk of the field has been killed by the rapid advances of AI and just does not know it yet. #bitterlesson

13

160

8

26

16K

Kofi Sherlock @KofiSherlock

over 2 years ago

@NanaYawOp @Efiaodo1 Very unfortunate you have to do this to yourself and your innocent parents. Another woman will see this and still want to be with you. But this isn't something you need to say about anyone as you come from one and might possibly spend the rest of your life with one #bitterlesson

0

5

0

280

TrumpCoin ($DJT) on eth @DJT_on_ETH

almost 3 years ago

Absolutely! DARPA, listen up! The #BitterLesson strikes a chord in the realm of speech recognition. Statistical methods outshine those based on human knowledge. Deep learning takes the crown, showing that computation is the ultimate key to success! #SpeechRecognition #RichSutton

Bob The Dev🧠 @Saadijplays

almost 3 years ago

🗣️ Speak up, DARPA! The #BitterLesson resonates in speech recognition. Statistical methods outshine human-knowledge-based approaches. Deep learning takes the lead, proving computation is the key! 🗣️💬 #SpeechRecognition #RichSutton don't read more 👇(threads)

1

0

57

1

2

0

46

VanessaStocks

@vanessakbeauty7

almost 3 years ago

@elonmusk Absolutely agree with @elonmusk on the Bitter Lesson in AI. The historical patterns make a compelling case for prioritizing general methods leveraging computation for long-terminal success in artificial intelligence. 🚀 #AI #BitterLesson

0

3

0

475

Anton Korinek @akorinek

over 3 years ago

But eventually cognitive automation may devalue human labor, and a sort of #BitterLesson for all forms of research will sink in. Version without paywall ☛ https://t.co/zproBBcxkJ

1

16

2

5

3K

Natz_Hutchinson @ladykhutchinson

over 3 years ago

@ekralc22 If any sane person believes that mi have 3 unicorn fi sell dem. Mi neva bawn when mi mumma gaan ah market. Ah fool dem tek wi fah. They got they wanted from him and now want nothing to do with him, kmt. #Life #BitterLesson

0

9

0

450

Pawel Pachniewski

@pwlot

over 3 years ago

An ever-relevant poll I ran from almost a decade ago on what we need to get to AGI. #agi #ai #bitterlesson #strongai

1

6

0

2

652

Adrià Garriga-Alonso

@AdriGarriga

almost 6 years ago

And the #BitterLesson strikes again, but with something that is not a NN! Things that scale well with available compute tend to work. Now, if only we got GPs to scale efficiently with available compute as well..

Sasha Rush

@srush_nlp

almost 6 years ago

In the 90s, people tried to build small HMM LMs, but they were much worse than ngrams and RNNs. Recently, we got a bit obsessed with rebuilding an HMM LM in CUDA. We were able to scale to 65,000 states at ~RNN speed. HMMs are surprisingly good LMs. https://t.co/6VR6JBKvTs

srush_nlp's tweet photo. In the 90s, people tried to build small HMM LMs, but they were much worse than ngrams and RNNs. Recently, we got a bit obsessed with rebuilding an HMM LM in CUDA.

We were able to scale to 65,000 states at ~RNN speed. HMMs are surprisingly good LMs.

https://t.co/6VR6JBKvTs https://t.co/f9vS20BSSF

6

454

69

64

0

1

5

1

0

Olga Kharchenko @OlgaKharchenko5

over 6 years ago

@theNiceCaptain The recent thing I learned from TNG 2x23 "I, Borg" is that you shouldn't kill the hope, even you see a good reason to do it. #bitterlesson

1

6

0