Explore intelligence

Report · Money

Small models trained on proprietary data can beat frontier models on narrow tasks

Armstrong gives the example that a small open-weight model trained on 100,000 compliance cases within Coinbase's proprietary data can outperform a frontier model on a compliance review queue. He also describes an emerging agentic marketplace of specialist agents, such as one that is just really good at ordering pizza or an incredible designer trained on tons of proprietary Figma data, and notes much sub-30-cent agentic commerce today is information gathering, like venture firms compiling paywalled data or recruiters scraping LinkedIn.

No Priors: Artificial Intelligence | Technology | Startups · Coinbase’s Everything Exchange: Agentic Finance, Stablecoins, and Tokenization with CEO Brian Armstrong

Claim from Brian Armstrong

Specialist-agent marketplace is speculative; the compliance example is publisher-reported from the speaker.

Report · Money

Crypto powers only ~0.5% of global GDP today

Armstrong states crypto currently powers only about 0.5% of global GDP, with roughly 700 million people having used or holding crypto but only 50-100 million using it monthly — so the mission is far from finished, and getting adoption to a billion or more users is how the financial system gets truly updated, achievable with humans and with agents.

No Priors: Artificial Intelligence | Technology | Startups · Coinbase’s Everything Exchange: Agentic Finance, Stablecoins, and Tokenization with CEO Brian Armstrong

Claim from Brian Armstrong

medium

Recommendation · Money

Don't run multiple companies at once too early

Armstrong advises not trying to run multiple companies at once too early — the distraction is dangerous and arguably irresponsible — and to limit concurrent ventures to a small handful so that simultaneous failures don't destroy your sleep and family life.

No Priors: Artificial Intelligence | Technology | Startups · Coinbase’s Everything Exchange: Agentic Finance, Stablecoins, and Tokenization with CEO Brian Armstrong

Recommendation from Brian Armstrong

Recommendation · Money

Give each employee a personal company 'brain' they can take with them

Armstrong recommends giving each employee a personal company 'brain' they can take with them when they leave — Coinbase is creating policies to enable this.

No Priors: Artificial Intelligence | Technology | Startups · Coinbase’s Everything Exchange: Agentic Finance, Stablecoins, and Tokenization with CEO Brian Armstrong

Recommendation from Brian Armstrong

Recommendation · Money

Employees' AI 'brains' should be portable when they leave the company

Armstrong believes individual employees' AI brains should be portable — they should be able to take them with them when they leave Coinbase — and Coinbase is creating policies to that effect.

No Priors: Artificial Intelligence | Technology | Startups · Coinbase’s Everything Exchange: Agentic Finance, Stablecoins, and Tokenization with CEO Brian Armstrong

Recommendation from Brian Armstrong

Report · Money

Scaling laws in three dimensions: perturbations, batch size, and sharding

Scaling laws exist in three dimensions for this training approach: number of perturbations, batch size, and degree of model sharding. Increasing batch size reduces gradient estimate noise — a previously known scaling dimension. The newly discovered dimension: sharding the model more and more continually improves results, better and better.

Y Combinator · What If We Stopped Using GPUs? | YC Paper Club

Claim from Y Combinator host

Creator perspective · Money

Why not just use neurons? Computers' advantages over the brain

A questioner argues that if the goal of neuromorphics is to mimic the brain, the best approximation might be to just use the brain itself — i.e., use neurons. In response: there are things computers do way better than the brain, e.g., computer clock speed is much faster, and computers offer effectively infinite storage capacity and better retrieval, whereas the brain has biological limitations in memory. Using biological neurons for computation is possible in principle, but the open question is whether it can scale; it's a direction of research not really gone down yet.

Y Combinator · What If We Stopped Using GPUs? | YC Paper Club

Claim from Alok

partial

Creator perspective · Money

Backprop would not work in alternative substrates like optics or neuromorphics

Backprop would not work in any of the alternative substrates (optics, neural compute, neuromorphics), partly because storing bits in light is very expensive and hard. In optical computing the cost per gigaflop is way, way low — computation in light is like free; you can do lots of it, you just can't easily store stuff. There are also non-differentiability and activation issues, and VRAM is very expensive. Bandwidth with light is pretty much instantaneous; with neural compute, even a copper wire could probably do it.

Y Combinator · What If We Stopped Using GPUs? | YC Paper Club

Claim from Y Combinator host

Report · Money

Scaling biocomputing: two open problems — frontier intelligence on cells and serving it to billions

Biocomputing has two main problems: getting frontier levels of intelligence on the cells, and figuring out how to serve this intelligence to billions of people in the world. We kind of know how to get intelligence on cells because we've seen it — we're all intelligent beings on the same substrate. What we haven't seen is distributed computing on cells; we fundamentally don't know how to do this yet. On scalability, the honest answer is 'we're trying to find out,' and the team has assembled a strong team toward scaling the technology. The team never truly measured the sample efficiency of the system down to the wire.

Y Combinator · What If We Stopped Using GPUs? | YC Paper Club

Claim from Y Combinator

Explicitly framed as unknown/open questions by the speaker.

Report · Money

PPO critic used to keep feedback to cells stable

The PPO critic was used to keep feedback to the cells stable, providing information over time about how well the cells are doing.

Y Combinator · What If We Stopped Using GPUs? | YC Paper Club

Claim from guest

Creator perspective · Money

Why this loses to backprop today: expensive forward passes — unless compute is optical or neuromorphic

The reason this approach loses to backprop is that the forward pass remains very expensive in conventional (electronic) compute. If forward passes were cheap — in optical computing they are extremely cheap, just beams of light doing computation, effectively free — perturbation-based training could become competitive. In neurons (neuromorphic computing), forward passes run at about 20 watts.

Y Combinator · What If We Stopped Using GPUs? | YC Paper Club

Claim from Y Combinator host

hypothetical-conditional

Creator perspective · Money

The hard problems for optical computing: conversion costs, calibration energy, and nonlinearities

A major challenge: data and models live in digital electronic memories, and converting digital information to analog and then optical domain (and back) carries a large energy price — both on the way in and on the way out. Programming and keeping optical devices calibrated and precise also requires actively spending energy. And while passive optics supports linear operations at massive scales, nonlinear activation functions require additional mechanisms and are not as easily accessible as with transistors or digital electronics — a bigger challenge for AI workloads. One attempted solution used short pulses of very intense light in multimode fibers, where light-matter interactions create optical nonlinearities — but this adds another layer of complexity.

Y Combinator · What If We Stopped Using GPUs? | YC Paper Club

Claim from Ilker (optical computing expert)

Principle · Money

Fixed-weight approach chosen because optical weights are expensive to rewrite

The reason the team took an application-specific and fixed-weight approach is that it is very expensive to rewrite or keep weights in the optical domain.

Y Combinator · What If We Stopped Using GPUs? | YC Paper Club

Claim from Ilker (narrator)

Report · Money

Research directions for neuromorphic edge hardware

Edge chips that need to sense and react and adapt can take inspiration from the brain's continuous sensing, e.g., a sensor on a drone doing something in real time. Researchers are exploring spiking networks (training a neural net on spikes rather than conventionally), on-device learning (can the network itself self-improve based on what it perceives?), and event-driven computation in the same way the brain works. Intel has done the most work on neuromorphic edge hardware, including a chip from its R&D lab and a drone that can fly itself with it.

Y Combinator · What If We Stopped Using GPUs? | YC Paper Club

Claim from Alok

Creator perspective · Money

Perturbation-based training doesn't scale to big models — unless you shard

In zeroth-order/forward-pass-only training, the error in the gradient estimate scales worse (linearly) as model size grows, becoming less and less accurate. You cannot train big models with this: a 10-billion-parameter model would require an extremely large, extremely infeasible number of perturbations — certainly more flops than a forward and backward pass. But if you shard the model, the gradient noise is capped by the size of each expert: only about 64 perturbations per training step (or fewer) are needed and it works completely fine.

Y Combinator · What If We Stopped Using GPUs? | YC Paper Club

Claim from Y Combinator host

Report · Money

Low-value agentic transactions are mostly information gathering

Most low-value agentic transactions are information gathering: venture firms compiling paywalled data, recruiters scraping LinkedIn, and agents calling other specialized agents for source data.

No Priors: Artificial Intelligence | Technology | Startups · Coinbase’s Everything Exchange: Agentic Finance, Stablecoins, and Tokenization with CEO Brian Armstrong

Claim from Brian Armstrong

Principle · Money

AI agents will outnumber humans; humans still need to advance

Armstrong believes there's going to be probably more AI agents than humans in the not-too-distant future, but humans still need to advance and progress. He also describes the hybrid of biological and digital intelligence with brain-machine interfaces, where some people will choose to merge and some will not.

No Priors: Artificial Intelligence | Technology | Startups · Coinbase’s Everything Exchange: Agentic Finance, Stablecoins, and Tokenization with CEO Brian Armstrong

Claim from Brian Armstrong

moderate

Creator perspective · Money

Debate: will gene editing diminish or increase human diversity?

The host wonders whether, in a future where gene editing is common, there may be packages you buy — the intelligence package, the health package — and whether these will diminish the neurodiversity of humanity by pushing everybody down the same package, like buying the Ford Model T of intelligence. He notes there are five or six different ways to imagine different forms of intelligence — processing speed, recall, etc. — and wonders whether parents will default opt into the basic package. Armstrong responds that he actually thinks gene editing could increase diversity, not decrease it, because there's just unlimited diversity of human preferences: parents would each choose differently — a great artist, a great general, a great engineer, a great politician. The host counters that this assume

No Priors: Artificial Intelligence | Technology | Startups · Coinbase’s Everything Exchange: Agentic Finance, Stablecoins, and Tokenization with CEO Brian Armstrong

Claim from Brian Armstrong and host

moderate

Report · Money

Coinbase's AI adoption: coding tools, ML for risk and fraud, and AI-routed customer support

Coinbase started AI adoption with the basics — coding tools and hooking AI into its data repositories — has long used machine learning for risk prevention and fraud prevention, and now routes customer support inquiries through AI.

No Priors: Artificial Intelligence | Technology | Startups · Coinbase’s Everything Exchange: Agentic Finance, Stablecoins, and Tokenization with CEO Brian Armstrong

Claim from Brian Armstrong

Principle · Money

Aging as the root of disease and the trillion-dollar skin rejuvenation market

Armstrong explains that most of the diseases that take us out are highly correlated with age — in your 20s your body recovers from errant cancerous cells or liver damage, but as you age, cells lose this function. The host explains that the Amanaka factors were able to change the imprint for both age as well as cell type, and Armstrong is focusing on just the age component — reversing the potency of the cell to something with more regenerative capabilities. Armstrong estimates that skin rejuvenation alone is probably a trillion-dollar market.

No Priors: Artificial Intelligence | Technology | Startups · Coinbase’s Everything Exchange: Agentic Finance, Stablecoins, and Tokenization with CEO Brian Armstrong

Claim from Brian Armstrong

moderate