Aug 30, 2024 · 37m · no-priors
No Priors Ep. 79 | With Magic.dev CEO and Co-Founder Eric Steinberger
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Magic co-founder and CEO Eric Steinberger discusses his company's mission to achieve AGI through code-centric models, ultra-long context windows, and autonomous software engineering colleagues. He also explores the architectural tradeoffs of inference-time compute and the societal imperative of safely navigating a bimodal post-AGI future.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 15.8% of the talking time here. How this is scored →
The hosts as informed peer 5.6 Guest teaching 4.6 Guest disagreement 1.9 The hosts pushing back 2.3
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 32:11 Rejecting the frontier math problem premise
Eric explicitly pushes back against Sarah's question about choosing a grand open problem like Riemann, insisting that focusing on specific questions misses the existential point of steering the bimodal distribution.
Hardest push from the hosts ▶ 15:17 Sarah confronts Eric on previous compute estimatesSarah brings up their prior disagreement over whether Magic could get away with less compute than general labs, directly challenging his previous thesis.
Biggest teaching moment ▶ 6:35 Long-context learning vs retrieval through Sutton's Bitter LessonEric provides a foundational technical rationale for why massive context windows subsume retrieval heuristics by referencing Sutton's Bitter Lesson.
The host holds their own ▶ 30:54 Sarah cites 'The Wealth of Humans' on post-work identitySarah demonstrates deep domain familiarity by introducing Ryan Avent's 2016 work to elevate the discussion beyond standard UBI talking points into structural human purpose.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| The Inception of Magic and Code-Centric AGI | 5 | 6 | 1 | 1 | Elad demonstrates knowledge of Magic's early pioneering of large 5-million-token context windows. Eric explains the theoretical foundation by invoking Richard Sutton's Bitter Lesson and explaining in-context learning as an online optimizer rather than heuristic retrieval. | |
| Balancing Training and Inference-Time Compute | 6 | 5 | 1 | 1 | Sarah and Elad actively participate in framing inference-time search, with Elad offering an analogy about human thinking pauses. Eric expands on the economic trade-off between training compute and test-time compute with the Terence Tao analogy. | |
| Recursive Self-Improvement, Alignment, and Automation | 4 | 4 | 2 | 1 | Elad prompts Eric on the recursive self-improvement roadmap. Eric dismisses short-term panic about safety while highlighting existential evolutionary risk, arguing recursive automation is the only controllable safety mechanism. | |
| Compute Scaling, Massive Clusters, and Productization Journey | 7 | 3 | 2 | 5 | Sarah presses Eric on his previous claims regarding lower compute requirements compared to foundation labs. Eric openly concedes that Sarah was right in their prior 1-on-1 debate and reveals Magic's massive cluster buildout. | |
| The Reliability Standard and Team Culture at Magic | 5 | 4 | 2 | 3 | Sarah drills down into the exact evaluation standard for developer tools, asking if the bar is skipping code review. Eric explains the step-function nature of software trust and his hiring philosophy for overlooked engineering talent. | |
| Societal Implications and Human Meaning in Post-AGI World | 7 | 4 | 2 | 3 | Hosts and guest engage in an intellectual exchange on post-AGI economics; Sarah cites Ryan Avent's 'The Wealth of Humans' to dispute simple UBI fixes, while Elad points to societal fragility resulting from abundance. | |
| Eric's Ultimate North Star and Bimodal Outcomes | 5 | 6 | 3 | 2 | Sarah asks Eric to name a specific grand challenge (like Riemann or Navier-Stokes) he wants Magic to solve. Eric politely rejects the premise, explaining that answering specific math problems is a trivial side effect compared to steering clear of catastrophic bimodal risks. |
Statements from this episode (19)
Steinberger: AGI can be reduced to an AI that writes code and tests ideas
“If you your end goal is to have a system that can do everything you can reduce that to building a system that can build that system. And so that minimal system is a system that writes code and comes up with ideas and can validate those by writing code and runn…”
Steinberger: In-context learning functions as an online optimizer
“I think of that as some sort of a, as an online optimizer in a sense that instead of compressing a set of data, you're trying to learn an optimizer.”
Steinberger: Long context windows are fundamentally superior to RAG
“Retrieval selects a subset of data for one completion. Our model sees all the data all the time. Clearly, a subset of data for the whole completion is a subset of all the data all the time. So if retrieval was optimal, our system could learn it. And it just tu…”
Steinberger: AI performance requires trading off training compute against inference compute
“Well, so you can think of model performance as some function of training compute times some function of inference time compute. Now those are specific functions that are just scaling law things that you can like model, but the general Way to think about it is …”
Steinberger: AI inference spending will follow an extreme power-law distribution
“There'll be users who will want to spend less money, and users will want to spend more money, and this is likely going to follow some sort of very sort of spiky distribution, where there'll be like four users spending a million dollars, and, you know, four bil…”
Steinberger: Pure training compute will fail to solve frontier intellectual problems
“Even the best mathematicians in the world, for the frontier of mathematics, require a long time to solve the problem. So, so, so, like, I would love to have a Terence Tao in my computer, but I would then still need to run Terence Tao for a year of human thinki…”
Steinberger: AI alignment is only solvable via recursive automated models
“The only way to sort of reasonably approach this is to iteratively ask your model to solve alignment and safety at that stage, not, not, you know, surely you can also ask it to solve your product level problems, but like that, that's nice, but that's not the f…”
Steinberger: Magic is announcing one of the largest compute clusters ever built
“Also in parallel with the release of this, we're going to be announcing what is going to be one of the largest clusters to ever be built.”
Steinberger: 99% of users wouldn't notice Gemini 4 replacing 1.5 Pro
“I don't think, if you took Gemini 1.5 pro, and you put it into Google, and you, like, did all the, you know, fine tuning properly, and then you swap it with Gemini four, I don't think anyone would notice, unless, like, you go, like, prove Riemann, right? But, …”
Steinberger: Magic skipped launching code completions because next-gen models will obsolete them
“Like, we decided not to launch completions, because it's just obviously going to get killed by the next thing.”
Steinberger: AI coding tool adoption will be a bimodal step-function
“I just genuinely think that there is a gigantic market that gets unlocked in a step function moment where users decide that they're no longer going to use VS code to write code and send it to their colleagues. They're going to use magic or whoever ends up doin…”
Steinberger: General AI coding models will span software verticals without specialization
“No, I think if you can, the leap to doing all use cases is small. Like you can build a UI builder and then it's like a normal UI builder, or you can build a true great UI builder driven by AI with some added features for that vertical. But then you can do the …”
Steinberger: Magic will never poach frontier lab staff for training IP
“We're not buying the IP by poaching someone from like a lab who tells us how they train GPT. Never done this. Will not do it.”
Steinberger: Capitalism and competition are the only mechanism to navigate AGI safely
“The reality is I really truly believe that capitalism and competition are the only chance we have to provide a, an optimizer that is capable of getting us to the right place. I think we need the right guardrails to do that.”
Steinberger: Etsy exemplifies massive future demand for post-AGI human-made goods
“Etsy is a great proof of concept for what happens after AGI. This is completely useless. Like you could just buy the made in China product. It looks the same, but it's not made by the human, you know, so. That's the thing, I think. Like, that will grow really …”
Steinberger: It will eventually be responsible to replace human CEOs with AI
“There'll be some CEO system, and then, like, the responsible decision will be to, you know, like, have that thing be the CEO, and then, so that will happen at some point.”
Steinberger: A mediocre post-AGI future does not exist; outcomes are bimodal
“Because I can't come up with, like, a mediocre AGI future. It doesn't exist.”
Steinberger: Software tools will optimize for AI agents like websites for Google
“The way websites had optimizations made for Google search crawlers crawling, I think there will be tool optimizations made for AI, and for those that don't have it, the models will just use it natively”
Steinberger: AGI Companies Will Swallow AI Wrapper Startups
“All, you know, this army of wrapper companies is going to get swallowed by AGI companies doing their own agent stuff. The same thing is going to happen there. Like, if you build your own tools, it's just going to get swallowed by simply the model learning to a…”