The Man Who Might SOLVE AI Alignment — Dr. Steven Byrnes, AGI Safety Researcher @ Astera Institute

36 snips

Aug 1, 2025

Dr. Steven Byrnes, an AI safety researcher at the Astera Institute and a former physics postdoc at Harvard, shares his cutting-edge insights on AI alignment. He discusses his 90% probability of AI doom while arguing that true threats stem from future brain-like AGI rather than current LLMs. Byrnes explores the brain's dual subsystems and their influences on decision-making, emphasizing the necessity of integrating neuroscience into AI safety research. He critiques existing alignment approaches, warning of the risks posed by misaligned AI and the complexities surrounding human-AI interaction.

Ask episode

AI Snips

Chapters

Books

Transcript

Episode notes

INSIGHT

Two Main Brain Subsystems

The brain consists of two major subsystems: the steering brain and the learning from scratch brain.
The steering brain encodes innate business logic, while the learning brain learns flexibly during life through algorithms.

INSIGHT

Cerebellum as Predictive LLM

The cerebellum acts like a large predictive learning algorithm, akin to an LLM focusing on prediction.
It solves latency issues in the brain by anticipating signals 0.2 seconds ahead, improving motor control and other processes.

INSIGHT

LLMs' Limitations with Complex Knowledge

Large language models (LLMs) are powerful but have intrinsic limitations, especially with complex, interconnected knowledge outside their training data.
They struggle with long-term planning involving idiosyncratic real-world details humans can learn during life.

Get the Snipd Podcast app to discover more snips from this episode

Get the app

Dr. Steven Byrnes, UC Berkeley physics PhD and Harvard physics postdoc, is an AI safety researcher at the Astera Institute and one of the most rigorous thinkers working on the technical AI alignment problem.

Steve has a whopping 90% P(Doom), but unlike most AI safety researchers who focus on current LLMs, he argues that LLMs will plateau before becoming truly dangerous, and the real threat will come from next-generation "brain-like AGI" based on actor-critic reinforcement learning.

For the last five years, he's been diving deep into neuroscience to reverse engineer how human brains actually work, and how to use that knowledge to solve the technical AI alignment problem. He's one of the few people who both understands why alignment is hard and is taking a serious technical shot at solving it.

We cover his "two subsystems" model of the brain, why current AI safety approaches miss the mark, his disagreements with social evolution approaches, and why understanding human neuroscience matters for building aligned AGI.

* 00:00:00 - Cold Open: Solving the technical alignment problem

* 00:00:26 - Introducing Dr. Steven Byrnes and his impressive background

* 00:01:59 - Steve's unique mental strengths

* 00:04:08 - The cold fusion research story demonstrating Steve's approach

* 00:06:18 - How Steve got interested in neuroscience through Jeff Hawkins

* 00:08:18 - Jeff Hawkins' cortical uniformity theory and brain vs deep learning

* 00:11:45 - When Steve first encountered Eliezer's sequences and became AGI-pilled

* 00:15:11 - Steve's research direction: reverse engineering human social instincts

* 00:21:47 - Four visions of alignment success and Steve's preferred approach

* 00:29:00 - The two brain subsystems model: steering brain vs learning brain

* 00:35:30 - Brain volume breakdown and the learning vs steering distinction

* 00:38:43 - Cerebellum as the "LLM" of the brain doing predictive learning

* 00:46:44 - Language acquisition: Chomsky vs learning algorithms debate

* 00:54:13 - What LLMs fundamentally can't do: complex context limitations

* 01:07:17 - Hypothalamus and brainstem doing more than just homeostasis

* 01:13:45 - Why morality might just be another hypothalamus cell group

* 01:18:00 - Human social instincts as model-based reinforcement learning

* 01:22:47 - Actor-critic reinforcement learning mapped to brain regions

* 01:29:33 - Timeline predictions: when brain-like AGI might arrive

* 01:38:28 - Why humans still beat AI on strategic planning and domain expertise

* 01:47:27 - Inner vs outer alignment: cocaine example and reward prediction

* 01:55:13 - Why legible Python code beats learned reward models

* 02:00:45 - Outcome pumps, instrumental convergence, and the Stalin analogy

* 02:11:48 - What’s Your P(Doom)™

* 02:16:45 - Massive headroom above human intelligence

* 02:20:45 - Can AI take over without physical actuators? (Yes)

* 02:26:18 - Steve's bold claim: 30 person-years from proto-AGI to superintelligence

* 02:32:17 - Why overhang makes the transition incredibly dangerous

* 02:35:00 - Social evolution as alignment solution: why it won't work

* 02:46:47 - Steve's research program: legible reward functions vs RLHF

* 02:59:52 - AI policy discussion: why Steven is skeptical of pause AI

* 03:05:51 - Lightning round: offense vs defense, P(simulation), AI unemployment

* 03:12:42 - Thanking Steve and wrapping up the conversation

* 03:13:30 - Liron's outro: Supporting the show and upcoming episodes with Vitalik and Eliezer

Show Notes

* Steven Byrnes' Website & Research — https://sjbyrnes.com/

* Steve’s Twitter — https://x.com/steve47285

* Astera Institute — https://astera.org/

Steve’s Sequences

* Intro to Brain-Like-AGI Safety — https://www.alignmentforum.org/s/HzcM2dkCq7fwXBej8

* Foom & Doom 1: “Brain in a box in a basement” — https://www.alignmentforum.org/posts/yew6zFWAKG4AGs3Wk/foom-and-doom-1-brain-in-a-box-in-a-basement

* Foom & Doom 2: Technical alignment is hard — https://www.alignmentforum.org/posts/bnnKGSCHJghAvqPjS/foom-and-doom-2-technical-alignment-is-hard

---

Doom Debates’ Mission is to raise mainstream awareness of imminent extinction from AGI and build the social infrastructure for high-quality debate.

Support the mission by subscribing to my Substack at DoomDebates.com and to youtube.com/@DoomDebates

Get full access to Doom Debates at lironshapira.substack.com/subscribe