Aji Ghose (@AI_Modeller) header
Aji Ghose (@AI_Modeller) profile picture

Chief Scientist (AI Research & NLP) @ChattermillAI | AI/ML, Semantics | 🤖 PhD Computational Cognitive Science | MSc Computer Science

London, England
Joined March 2016
karpathy says we’re a decade away from AGI, because we don’t yet know how to make systems learn continuously. the deeper problem is that we’ve built this entire field on metaphors, not mechanics. we keep saying AI can think, reason, remember, create. but those are human verbs, not model capabilities. AI isn’t intelligent. it’s efficient. it doesn’t reason . it pattern-matches. it doesn’t remember. it reconstructs. it doesn’t reflect. it re-runs. we confuse language with understanding. just because a model can describe thought doesn’t mean it’s having one. real intelligence has intent. it knows why it’s thinking. AI predicts what comes next. and yet, even without intent, these systems are starting to functionally mimic cognition. they reason, recall, and reflect. not consciously, but effectively. that’s why both statements can be true. AI is a bubble. because capital, hype, and valuations have outpaced genuine capability. but it’s also here to stay. because the direction of progress is right. the crash will clear the noise. what remains will be systems that truly learn. memory that compounds, feedback that refines, intelligence that grows by living inside workflows. we’ll look back on this phase the way we look at the early web: messy and magical. the beginning of machines that finally learn, not just perform.
The @karpathy interview 0:00:00 – AGI is still a decade away 0:30:33 – LLM cognitive deficits 0:40:53 – RL is terrible 0:50:26 – How do humans learn? 1:07:13 – AGI will blend into 2% GDP growth 1:18:24 – ASI 1:33:38 – Evolution of intelligence & culture 1:43:43 - Why self driving took so long 1:57:08 - Future of education Look up Dwarkesh Podcast on YouTube, Apple Podcasts, Spotify, etc. Enjoy!
110
192
36
1,630
264,469
💯
Everyone posting about the Dwarkesh interview (including Dwarkesh himself!) is missing this subtle point. When LLMs imitate, they imitate the ACTION (ie the token prediction to produce the sequence). When humans imitate, they imitate the OUTPUT but must discover the action
27
46
5
650
101,514
Dwarkesh and I had a frank exchange of views. I hope we moved the conversation forward. Dwarkesh is a true gentleman.
.@RichardSSutton, father of reinforcement learning, doesn’t think LLMs are bitter-lesson-pilled. My steel man of Richard’s position: we need some new architecture to enable continual (on-the-job) learning. And if we have continual learning, we don't need a special training phase - the agent just learns on-the-fly - like all humans, and indeed, like all animals. This new paradigm will render our current approach with LLMs obsolete. I did my best to represent the view that LLMs will function as the foundation on which this experiential learning can happen. Some sparks flew. 0:00:00 – Are LLMs a dead-end? 0:13:51 – Do humans do imitation learning? 0:23:57 – The Era of Experience 0:34:25 – Current architectures generalize poorly out of distribution 0:42:17 – Surprises in the AI field 0:47:28 – Will The Bitter Lesson still apply after AGI? 0:54:35 – Succession to AI
76
197
42
3,545
657,348
Many people think CS is just DSA. It's not. My CS curriculum had: 0. Data Structures and Algorithms 1. Calculus 2. Linear Algebra 3. Physics 4. Mathematical Physics 5. Differential Equations 6. Theory of Functions of a Complex Variable 7. Theory of Probability and Mathematical Statistics 8. Calculus of Approximations 9. Functional analysis 10. Numerical Methods 11. Optimisation Methods 12. Cryptography 13. Compression Algorithms 14. Game Theory 15. Algebra and Geometry 16. Discrete Mathematics 17. Automata-based Programming 18. Formal Language Theory 19. Compilers and Interpreters 20. Computational Complexity Theory 21. Math Logic 22. Type Theory 23. Lambda Calculus 24. Functional Programming 25. OOP Design Pattern 26. Machine Learning 27. AI 28. Hardware Architecture 29. Java 30. C++ 31. C 32. Computational Geometry 33. Operating Systems 34. Computer Networks 35. Databases, Relational Algebra, and SQL 36. Parallel Programming 37. Economics 38. Organisation and management of business processes 39. History 40. Philosophy 41. English 42. Health and Safety Training Course 43. PE
113
298
15
3,399
219,070
Aji Ghose retweeted
hypothetical two-year degree in AI: - Coding in Python - Semiconductors 101 - Intro to machine learning - Intro to data science - Data visualization & dimensionality reduction - Machine learning engineering - Language models - Deep learning (basics & architectures) - Reinforcement learning - Computer vision - Generative modeling - Robotics & Planning - LLMs 1 (Pre-training) - LLMs 2 (Post-training) - GPU Architecture & Intro to CUDA - Energy & Datacenters - AI Governance - AI Safety - Federated & private learning - Advanced CUDA
51
93
5
968
68,750
The Neocognitron was trainable, but not via backpropagation. It used analogue threshold units, not ReLU, sigmoid, or tanh. CNNs became scalable and trainable with backpropagation, notably through LeNet.
1
1
11
4,770
Replying to @SchmidhuberAI
True, but these CNNs were not trainable. AFAIK @ylecun is the one training CNNs for the first time.
3
3
37
12,769
Aji Ghose retweeted
The downfall is real 😬
10
25
4
205
22,845
Aji Ghose retweeted
If Claude is really doing so much of the coding for Anthropic, why haven't they used it to create a fucking ui for Claude Code? It's 2025. Why the fuck am I forced to use a cli for everything as if it were 1995?
598
166
79
6,836
1,033,276
Aji Ghose retweeted
"pip install" is dead. "uv add" is the new king.
180
219
51
3,534
512,613
Aji Ghose retweeted
Sergey Brin and Larry Page must be rolling in their graves now
google is entering its yahoo phase
273
423
88
12,049
1,608,923
Out last week in @cogsci_soc! @samhforbes and I ”urge extreme caution in the use of LLMs in classrooms lest we further normalize: pupils losing their privacy, reducing contact between learner and educator, deskilling teachers & polluting the environment”‼️ PDF/DOI links below.
3
21
1
67
3,458
🚨 A message in startup mahakumbh event at Bharat Mandapam, New Delhi.
618
3,092
346
23,515
1,318,663
Replying to @JeffDean
Yes I can; as can you. But I'm primarily interested in what's widely available in the community, where a single 4090 GPU machine is already a very rich investment. Remember also that 3090s were the last consumer card with nvlink, so 4090 and 5090 cards aren't good at multi gpu
7
2
213
15,043
Neural networks don’t have to be distilled into other neural networks. They can be distilled into decision trees or sets of rules. And then interpretability becomes dramatically easier.
48
91
21
1,314
146,258
Aji Ghose retweeted
Happy birthday Joseph Fourier, whose 1822 equation allows us to listen to mp3s today: bit.ly/22kbNfi
43
587
76
2,884
254,606
Our team at Microsoft Research India is looking for a Research Intern for a 6 month position. The position will be on-site in BLR. You will get to work on multilingual data, modelling and evals. Please DM me with a short blurb about yourself and your CV/Resume.
7
17
1
201
15,840
Can we enhance interoception by controlling the heart? Thrilled to share our new study, led by @AshleyTyrer, where we use computational modeling to show that blockading peripheral noradrenaline uniquely alters awareness of heart rate & breathing! biorxiv.org/content/10.1101/… 🧵👇
1
21
3
71
10,317
I am pretty happy with this 30-minute summary of my views on the current state of AI and alignment. youtube.com/watch?v=w177Ov-Y…
11
102
14
717
123,030
Aji Ghose retweeted
someone made a Lego Transformer
15
93
16
1,028
72,480