Uploaded January 2025 | Updated September 2026, 1 week ago
Machine Learning Street Talk
|
Schmidhubered
🔇 TAP TO UNMUTE
![Transformers Need Glasses! [Federico Barbero]](https://i.ytimg.com/vi/FAspMnu4Rt0/mqdefault.jpg "Transformers Need Glasses! [Federico Barbero]
SPONSOR MESSAGES:
***
CentML offers competitive pricing for GenAI model deployment, with flexible options to suit a wide range of models, from small to large-scale deployments. Check out their super fast DeepSeek R1 hosting!
https://centml.ai/pricing/
Federico Barbero (DeepMind/Oxford) is the lead author of Transformers Need Glasses!, a paper revealing fundamental architectural limitations in transformer-based language models. The conversation explores why LLMs fail at seemingly trivial tasks like copying the last token of a sequence or counting repeated elements, tracing these failures to representation collapse — where internal representations of distinct inputs converge below machine precision as context length grows. Federico connects these findings to information propagation theory from graph neural networks, showing how causal attention creates an inherent bias toward the start of a sequence, while training dynamics push models to attend to the most recent tokens, leaving the middle of long contexts as a blind spot. The discussion covers connections to spectral graph theory, heat equations on graphs, Petar Veličkovićs work on graph attention networks, the role of softmax in limiting sharp attention, and practical glasses — architectural tweaks that can help transformers see more clearly.
REFERENCES:
paper:
[00:01:05] Transformers Need Glasses!
https://proceedings.neurips.cc/paper_files/paper/2024/file/b1d35561c4a4a0e0b6012b2af531e149-Paper-Conference.pdf
[00:05:30] Softmax is Not Enough
https://arxiv.org/abs/2410.01104
[00:15:05] Graph Attention Networks
https://arxiv.org/abs/1710.10903
[00:43:23] Neural Networks and the Chomsky Hierarchy
https://arxiv.org/abs/2207.02098
[00:51:04] On the Measure of Intelligence
https://arxiv.org/abs/1911.01547
[01:00:35] Epistemic Foraging
https://www.frontiersin.org/journals/computational-neuroscience/articles/10.3389/fncom.2016.00056/full
LINKS:
Full Transcript: https://app.rescript.info/share/d43b3918dc7cb8c8b822ebef40b8f66f
Download PDF transcript: https://app.rescript.info/api/public/sessions/53f7d10ff5c72eca/pdf")
![MIGHT THE ROBOTS TAKE OVER? [Prof. Yoshua Bengio]](https://i.ytimg.com/vi/G1ARvwQntAU/mqdefault.jpg "MIGHT THE ROBOTS TAKE OVER? [Prof. Yoshua Bengio]
SPONSOR MESSAGES:
***
CentML offers competitive pricing for GenAI model deployment, with flexible options to suit a wide range of models, from small to large-scale deployments.
https://centml.ai/pricing/
Turing Award winner Professor Yoshua Bengio sits down with Tim Scarfe for a wide-ranging discussion on AI safety, the architecture of intelligence, and what happens when machines become as capable as our best researchers.
Bengio argues that the current race toward agentic AI systems carries fundamental risks that most people underestimate. He lays out how instrumental convergence and reward tampering could lead systems to develop goals misaligned with human interests not through malice, but through the basic logic of optimization. His proposed alternative: powerful AI tools that function as scientific oracles rather than autonomous agents, systems that can revolutionize medicine and science without needing goals of their own.
The conversation covers international AI governance and the game theory driving the US-China competition, the military implications of frontier AI capabilities, and why hardware-enabled governance may be the most promising path toward verifiable AI treaties. Bengio draws direct parallels to nuclear nonproliferation and warns that data centers will become strategic military assets.
On the technical side, Bengio discusses the return of recurrent architectures (the Were RNNs All We Needed? paper), his work on GFlowNets for probabilistic inference with discrete structures, a complexity-based theory of compositionality that tries to formalize what was previously just intuition, and the question of whether System 2 reasoning needs to be designed from first principles rather than bolted onto existing architectures.
Throughout, Bengio repeatedly emphasizes what he does not know a deliberate stance he contrasts with the overconfidence he sees driving dangerous policy decisions.
REFERENCES:
paper:
[00:00:15] AI Risk Statement
https://www.safe.ai/work/statement-on-ai-risk
[00:23:10] Reward Tampering and AI Safety
https://hdsr.mitpress.mit.edu/pub/w974bwb0
[00:44:30] Can a Bayesian Oracle Prevent Harm?
https://arxiv.org/abs/2408.05284
[00:52:00] Hardware-Enabled AI Governance Memo
https://yoshuabengio.org/wp-content/uploads/2024/08/FlexHEG-Memo_August-2024.pdf
[01:33:20] GFlowNet Foundations
https://arxiv.org/abs/2111.09266
[01:35:05] Complexity-Based Compositionality Theory
https://arxiv.org/abs/2410.14817
[01:37:50] Discrete Attractor States in Neural Systems
https://arxiv.org/abs/2302.06403
person:
[00:03:50] Professor Yoshua Bengio
https://yoshuabengio.org/
tool:
[00:40:45] Munk Debate on AI Existential Risk
https://munkdebates.com/debates/artificial-intelligence
[00:56:07] 2018 Turing Award
https://awards.acm.org/about/2018-turing
[01:12:35] EU AI Act Code of Practice
https://digital-strategy.ec.europa.eu/en/news/meet-chairs-leading-development-first-general-purpose-ai-code-practice
LINKS:
Full Transcript: https://app.rescript.info/share/1c4f010852d0db9a34dbeb043c0da26b
Download PDF transcript: https://app.rescript.info/api/public/sessions/baff3f8846202250/pdf
Yoshua Bengio:
https://x.com/Yoshua_Bengio
https://scholar.google.com/citations?user=kukA0LcAAAAJ&hl=en
https://yoshuabengio.org/
https://en.wikipedia.org/wiki/Yoshua_Bengio")
![Is Thermo AI the future? [Guillaume Verdon aka Beff Jezos]](https://i.ytimg.com/vi/HR-_U0Pzl1Y/mqdefault.jpg "Is Thermo AI the future? [Guillaume Verdon aka Beff Jezos]
Dr. Maxwell Ramstead hosts Guillaume Verdon physicist, founder of Extropic, and the mind behind the Beff Jezos persona and the Effective Accelerationism movement for a conversation that bridges thermodynamics, computing hardware, and civilizational philosophy.
Verdon traces his path from childhood fascination with theories of everything through theoretical physics at the Perimeter Institute, quantum computing at Google, and the founding of Extropic. The core technical insight: instead of fighting thermal noise at enormous energetic cost (as quantum computers do), thermodynamic computing harnesses it. Extropics chips use the natural stochastic physics of electrons to accelerate Markov chain Monte Carlo sampling the same class of algorithms that underpin diffusion models, energy-based models, and much of modern probabilistic ML.
The numbers are striking. Current wafer-scale AI systems consume 20+ kilowatts. A thermodynamic wafer with 1.5 billion p-bits and 20 billion parameters would run on 20 watts roughly what the human brain uses. Verdon argues this is not a coincidence: the brain is literally a thermodynamic computer operating near the Landauer limit, and Extropic is building silicon that works on the same principles.
The second half turns to philosophy. Verdon derives Effective Accelerationism directly from stochastic thermodynamics: thermodynamic selection pressure favors systems that capture and dissipate more free energy, so growth is not just desirable but physically inevitable. He and Ramstead explore hyperstition through the lens of active inference, the geopolitical stakes of the US-China technology race, and why Verdon views deceleration as a form of psychological warfare against Western competitiveness.
Recorded with Maxwell Ramstead hosting in place of Tim Scarfe.
***SPONSOR MESSAGE***
Google Gemini 2.5 Flash is a state-of-the-art language model in the Gemini app. Sign up at https://gemini.google.com
***
TIMESTAMPS:
00:00:00 Intro Montage & Sponsor
00:02:21 From Theories of Everything to Thermodynamic Computing
00:08:41 The Failure of Reductionism & Physics-Inspired AI
00:17:15 From Quantum to Thermodynamic Computing
00:23:56 How Thermodynamic Computers Actually Work
00:31:15 Moores Wall & The Brain as Proof of Concept
00:40:00 The Computing Stack of the Future
00:48:23 Why Current AI Will Cook Us to Death
00:50:38 Effective Accelerationism: The Philosophy of Beff Jezos
01:00:00 Hyperstition, Geopolitics & Avoiding Catastrophe
01:16:00 Closing: The Future of Extropic & Getting Involved
REFERENCES:
paper:
[00:04:40] It From Bit
https://cqi.inf.usi.ch/qic/wheeler.pdf
[00:06:15] Renormalization Group Theory
https://www.damtp.cam.ac.uk/user/dbs26/AQFT/Wilsonchap.pdf
[00:23:00] Metropolis-Hastings Algorithm
https://arxiv.org/abs/1504.01896
[00:40:00] Deep Information Bottleneck and Renormalization
https://arxiv.org/abs/1907.07331
[00:40:00] Energy-Based Models (EBMs)
https://www.researchgate.net/publication/200744586_A_tutorial_on_energy-based_learning
[01:03:00] Free Energy Principle and Active Inference
https://www.nature.com/articles/nrn2787
concept:
[00:05:30] Holographic Principle (AdS/CFT Correspondence)
https://en.wikipedia.org/wiki/Holographic_principle
[00:20:00] Markov Chain Monte Carlo (MCMC)
https://en.wikipedia.org/wiki/Markov_chain_Monte_Carlo
[00:21:10] Maxwells Demon and Information Theory
https://plato.stanford.edu/entries/information-entropy/
[00:29:45] Landauers Principle
https://en.wikipedia.org/wiki/Landauer%27s_principle
[01:11:40] Fishers Fundamental Theorem of Natural Selection
https://en.wikipedia.org/wiki/Fisher%27s_fundamental_theorem_of_natural_selection
LINKS:
Full Transcript: https://app.rescript.info/share/21936dd6830231c0194e8b2908f4d963
Download PDF transcript: https://app.rescript.info/api/public/sessions/a34e6a9a2ed827b9/pdf")
