Editorial illustration for Small team builds model in four months using the Eureka Machine to test ideas
Researchers Build AGI Model in 4 Months With Eureka
Small team builds model in four months using the Eureka Machine to test ideas
OpenAI says the AGI era has started. Nvidia's Jensen Huang goes further, calling it arrived. Yet some of the very people who built the models behind those claims are walking out the door.
Researchers who helped ship GPT-4, Gemini, and Llama are leaving OpenAI, Google DeepMind, and Meta to start their own labs, often with no product, no revenue, and nothing to show investors except a thesis. They're raising billions anyway.
THE DECODER's Frontier Radar series tries to make sense of moves like this six times a year, and issue #5 turns to the neolabs: 45 of them, sorted by research bet, with 27 profiled in depth and nearly 20 more covered briefly. Earlier issues tackled agentic AI, the measurable productivity gains from AI tools, the token economy behind generative models, and where the US-China AI race actually stands.
This issue asks a narrower question: if scaling today's recipe were enough to reach AGI, why are the people who know that recipe best betting that it isn't? One small team's approach, built fast and tested hard, shows what that bet looks like in practice.
What many founders share is a doubt. Not about whether much more powerful AI is coming; some think AGI is close. Their doubt is about whether today's recipe alone gets there: ever-larger language models, trained on internet data, refined with reinforcement learning.
Why this matters
A small team shipping a working model in four months, with an automated system proposing and testing most of the pipeline, is a bigger signal than the model itself. If Socher's account holds up, the bottleneck in frontier AI is shifting from headcount and compute hoarding to how well a lab can automate its own research loop. That's worth sitting with if you're at a startup trying to compete with labs that have ten times your budget: the Eureka Machine approach suggests the gap might be closing on the research side faster than on the compute side.
It also complicates the neolab story we've been tracking, dozens of ex-OpenAI, DeepMind, and Meta researchers raising billions with no product, betting the current recipe is tapped out. Socher's team just showed a cheaper, faster path using that same recipe, automated. We'd want to see the actual benchmarks and a repeat performance before calling this a pattern, but the claim itself should make every well-funded neolab explain why they need nine figures to do what a small team just did with a machine that tests its own ideas.
Related Reading
Research & Benchmarks Microsoft AI Chief: "If It Isn't Safe We Shouldn't Build It
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives LangSmith BYOC Expands to AWS, Azure, and Google Cloud
From Archives Alexa for Shopping adds text-based ordering and recommendations
Common Questions Answered
What is the Eureka Machine and how did it enable a small team to build a model in four months?
The Eureka Machine is an automated system that proposes and tests most of the research pipeline, allowing a small team to develop a working model in just four months. This approach represents a significant shift in frontier AI development, suggesting that automation of the research loop may be more valuable than simply increasing headcount and compute resources.
Why are researchers from OpenAI, Google DeepMind, and Meta leaving to start their own labs despite AGI being close?
Many departing researchers share a fundamental doubt about whether today's recipe of ever-larger language models trained on internet data and refined with reinforcement learning alone can achieve AGI. These founders believe that much more powerful AI is coming, but they want to explore alternative approaches rather than continue with the current scaling paradigm at major labs.
What does the shift from headcount and compute hoarding to research automation mean for AI startups?
The Eureka Machine approach suggests that the bottleneck in frontier AI is moving away from raw resources like compute and personnel toward how effectively a lab can automate its own research loop. This shift is significant for startups competing against labs with ten times their budget, as it indicates that superior research automation could potentially level the playing field.
Are investors funding these new AI labs despite them having no product or revenue?
Yes, investors are raising billions for these new labs founded by former researchers from major AI companies, even though many have no product, no revenue, and nothing to show except their research thesis and team credentials. This investment pattern reflects confidence in both the founders' expertise and their novel approaches to achieving more powerful AI systems.
Further Reading
- Recursive launches Flint, a laptop coding model it says was largely built by AI - TechCrunch
- Recursive shows first result of its AI-driven research process with Flint - The Decoder
- Frontier Radar #5: If AGI is so close, why is everyone leaving? The neolabs and their bets - The Decoder
- Introducing Flint - Recursive
