The 250 People Building Frontier AI Models

About 250 people are pushing frontier AI models forward every morning, and many now believe recursive self-improvement puts exponential intelligence one to two years away, according to Invest with the Best.

That belief was fringe a year ago. Now it is common inside the labs.

The full discussion can be found on Invest with the Best's YouTube channel.

The mood is violently competitive and globally stretched. Insecurity is rising because scale has changed what individual contribution feels like.

How frontier AI models change the attack surface

The issue is not a single demo. It is that powerful open-source models from the US, Europe and China are already available and in use everywhere, including recent poolside, Reflection, Nvidia and other releases.

When a frontier AI model can be downloaded and run locally, the attacker requirement drops. No API key, no rate limit, no remote monitoring.

Think of it like publishing the blueprint for a chemical plant instead of selling the chemical. Once the plant design is out, anyone can build a copy tuned to their own purpose.

The same capability that lets a defender scan code or assist biology research works for offensive cyber and bio work. Cost, sensitivity and latency push businesses to self-host, which broadens diffusion beyond what frontier labs can see or steer.

Researchers interviewed describe two disempowering views now circulating: the model will do the research anyway, or only compute scale matters. Both reduce human agency and understate risk.

What is not fixed and why it matters

For builders, cheap, controllable intelligence is coming, and businesses want it for economics and capacity. Restricting open models in the US would mostly hobble law-abiding American firms while adversaries ignore the rules.

What is not fixed is testing. There has been little rigorous safety evaluation for backdoor behavior in Chinese models or for bio and cyber offensive potential at the frontier. The right move is more measurement, not less openness.

Compute is the other gap. An infrastructure lead at a hyperscaler said nothing will move the needle at sufficient scale before 2030, citing natural gas and power constraints. Nuclear cost curves and data center buildouts are limited by permitting and alignment, not by technology.

That is why compute independence is now framed like energy independence. Every link from glass to fabs to GPUs has single points of failure, including the TSMC-centered supply chain. Projects like Jacob Helberg’s PACIA aim to map and fund alternate paths, and investors like Conviction are backing labor for data centers, nuclear and alternative chip architectures, but financing and physical build speed remain slow compared with software.

The interview also shows how early-stage decisions are made when intuition is scarce. Conviction’s Sarah Guo describes starting at an 8 or 9 on people and then hunting for holes, writing full memos even as a solo GP and sending them to partners like Mike, Bella, Pranov and outsiders such as Gokul or Dylan Field for pushback.

The 250-person focus and bets like Harvey for law or Sunday Robotics for general semi-humanoid systems illustrate a technology-forward filter: find tasks where next-token prediction maps cleanly to structured language or data, then back teams that treat data collection as a technical problem.

Gaps a knowledgeable reader still needs: which safety tests would actually detect backdoors or offensive uplift, what threshold triggers a block on open release, and what specific power and fab capacity would make the US independent by 2032.

Intelligence too cheap to meter will arrive. Whether it is broadly competitive depends on building the power and supply chain to run it.