Interview
Founder Eric Steinberger on Magic’s Counterintuitive Approach to Pursuing AGI
- Core Research Challenge: The remaining unsolved problem in AI is achieving general-domain, long-horizon reliability, which Eric Steinberger argues requires "inference time compute" or "test time compute."
- Analogy: Complex tasks like proving mathematical theorems, writing large software programs, or drafting complex essays require models to "think" about specific tokens rather than generating them token-by-token.
- Resource Allocation: The industry must shift toward spending "1 million times" the resources on productive reasoning for specific tokens rather than just 1x, 2x, or 10x.
- Eric Steinberger's Background & Education:
- Steinberger's obsession with AI began at age 14, driven by a desire for predictability and a resolve to avoid "decision paralysis" by automating tasks.
- He bypassed traditional academia by contacting DeepMind researcher Johannes (a student of David Silver) as a high schooler, proposing a 10-point plan to beat a specific algorithm.
- He collaborated with Noam Brown after Steinberger published "Single Deep Counterfactual Regret Minimization," which slightly outperformed Brown's "Deep Counterfactual Regret Minimization."
- Steinberger worked part-time at FAIR while a university student, noting Brown's ability to "pick the right problems" and grind on them consistently with high efficiency.
- He founded a non-profit, Climate Science, running it like a startup with ~1,000 volunteers and 20 engineers, but dropped out of university to focus on AI after realizing AGI was closer than 20 years.
- Philosophy on Research & Career:
- Mentorship Strategy: Steinberger advises sending detailed, 5-page emails to top researchers proposing specific plans to beat their algorithms; this "proactive" approach successfully secured mentorship from Brown and others.
- Learning Methodology: He emphasizes reading "every paper" to build a mental database for synthesis, noting that most great ideas are combinations of existing concepts (e.g., Transformers synthesizing attention and ResNet).
- Work Ethic: He advocates for "burning the boats" early with no backup plan, citing a formative moment where he persisted through doubt to finish a project after a mentor nearly gave up on him.
- Self-Awareness: He stresses being "really fucking honest" about weaknesses, as betting on strengths is futile if necessary skills (like engineering implementation) are missing.
- Magic's Mission & Strategy:
- Goal: Magic aims to build an AI software engineer capable of automating software engineering recursively, effectively allowing the company to "clone" its own workforce to improve alignment and model efficiency.
- Vertical Integration: Steinberger insists on training proprietary models rather than relying on APIs, arguing that the economic value of AGI will accrue at the hardware and model layer, not the application layer.
- Competitive Moat: He contends that while competitors have more funding, the "sufficient" compute required to build AGI is lower than the $100B+ revenue some companies generate; thus, capital is not the primary bottleneck.
- Team Structure: Magic operates with a small, high-velocity team (low tens of researchers) rather than scaling to hundreds, leveraging focused expertise on model training and long context.
- Agent Vision & Benchmarks:
- User Experience Goal: The target is a "colleague" tier agent where the user only defines the goal; the agent handles execution and clarification without requiring the user to micromanage or review code (targeting >99% reliability).
- Timeline: Steinberger predicts this level of competence is "very soon," estimating less than 10 years and likely "way less" than current predictions, noting that current math benchmarks are already surpassing market expectations.
- New Technical Announcement:
- Open Source Eval: Magic open-sourced a "hashless eval" for long-context testing to address the flaws of the "Needle in a Haystack" benchmark.
- Critique of Existing Methods: The traditional "Needle in a Haystack" test is flawed because models can rely on an implicit prior that the target is "weird" without actually memorizing the entire context.
- Methodology: Magic's new eval replaces semantic text with random noise (hashes) to force the model to process the entire context window, adding a "hop" dimension where the model must find multiple scattered hashes to verify retrieval capability.