Explore intelligence

Creator perspective · Money

ADC/DAC resolution and accuracy is a key problem in photonic accelerators

A key problem in photonic accelerators is ADC and DAC resolution/accuracy, which eventually affects prediction correctness over time in the operational infield. A speaker describes working on Enwise Light Matters on a silicon photonics accelerator, specifically on the DAC and ADC subsystems for the photonic core that would be the offload engine for most tensor ops. A project the speaker cannot discuss observed the ADC/DAC accuracy problem, which was the reason Lightmatter eventually pivoted to interconnects for HBMs and stopped work on their multispectral compute arm of research.

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

Claim from Guest speaker

described

Principle · Money

Perturbation-based training does not scale to big models

The speakers describe why perturbation-based optimizers lose to backprop: the error in the gradient estimate scales worse than linearly, becoming less accurate as model size grows, so you cannot train big models with it. Training a 10 billion parameter model would require an extremely large and infeasible number of perturbations, and the cost of forward pass remains really expensive. Sharding a model caps gradient noise to the size of each expert, requiring only about 64 perturbations per step for training to work fine.

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

Claim from Y Combinator Paper Club

scaling claims are the speakers' analysis

Creator perspective · Money

Spatial light modulator with mirror enabled prototyping and on-the-spot backpropagation

In their experimental implementation, the team used a spatial light modulator (a device generally used in light projectors) coupled with a mirror as an easier prototyping approach. By looking at camera images and input laser distributions, the team can do backpropagation on the spot and update weights, avoiding repeated cycles of microfabrication. The spatial light modulator setup gave the team a useful means for prototyping the optical system. For a useful optical compute unit in the future, the team would want to produce the layers via microfabrication and put them in place, which literature already shows is quite accessible.

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

Claim from Ilker

described

Report · Money

SPSA: reproducible noise makes zeroth-order learning work, but reproducibility is the hard part

With SPSA, if the noise is reproducible and premised off the same seed, it is completely learnable — two reproducible runs for forward difference, three for central difference — but the main issue is justifying how you get the same seed to a vector twice; zeroth-order methods may not be biologically plausible partly for this reason.

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

Claim from Alok

explicit

Creator perspective · Money

Entropy penalties instead of entropy bonuses in PPO

Traditionally in reinforcement learning you want a bit of an entropy bonus so you can start exploring the space around it. In their code they implemented entropy penalties, which is very abnormal for PPO. They actively penalized entropy because they found it was helpful and the cells provided sufficient entropy for exploration.

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

Claim from Y Combinator Paper Club speakers

The claim that penalizing entropy was helpful is the team's own finding, not an externally validated result.

Creator perspective · Money

Optical computing: passive weight masks, conversion costs, and nonlinear activation challenges

In optical computing, modulate input information on N mini modulators and let it propagate through a designed weight mask so weights modulate the light beam passively without spending extra active energy. But converting digital information to analog and then optical domain for optical processing requires paying a large energy price, and converting back after computation also costs significantly. Programming optical devices and keeping them calibrated and precise requires actively spending energy to put parameters or controls in place. Passive optics supports linear operations at massive scales but nonlinear activation functions require additional mechanisms not as easily accessible as in transistors or digital electronics. Using short pulses of very intense light in multimode fibers can

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

Claim from Ilker

The multimode-fiber nonlinearity approach is described as one of their tries, with added complexity acknowledged.

Report · Money

SOMA: cluster Common Crawl with TF-IDF/SVD, ship one parquet per GPU, train small experts, route at test time

SOMA works by taking the Common Crawl dataset, clustering it with TF-IDF and some SVD, then shipping one parquet file per GPU to N GPUs throughout the world with N clusters, and training a small expert (a small LCM, like 41K) on each. At test time you pull all the experts together and route with a little router — like the benefit of MoE, except it's a mixture of models, not just one layer of feed-forward layers. The premise is asking: how far can this go?

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

Claim from Y Combinator host

host's description; provenance flagged possible numeric error

Principle · Money

SPSA: reproducible noise makes optimization learnable

The host explains that with SPSA, if you can make the noise reproducible from the same seed, the optimization becomes completely learnable. He notes SPSA requires two forward evaluations for forward difference (one to perturb, one to scale the gradient) and three for central difference. A Stanford neuroscientist, Shraddha Druckman, is cited on the main obstacle to zero-order methods being biologically plausible: you cannot justify how you get the same seed to a vector twice.

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

Claim from Y Combinator Paper Club host

evaluation-count claim flagged as possible numeric error in the source

Report · Money

Toshi: Coinbase's internal agent harness that pays external vendors over X402

Coinbase built an internal agent harness called Toshi integrated with its tools; it can make calls to external vendors and pay using the agentic finance economy over the X402 protocols.

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

Claim from Brian Armstrong

Creator perspective · Money

Tokenized assets move wallet-to-wallet at the speed of information, like a WhatsApp message

With tokenized assets, money can move at the speed of information, just like a WhatsApp message — you can gift a share of stock to your nephew by sending it to their wallet, without going through some complex, antiquated stock transfer process. Tokenized assets can all be traded on the Everything Exchange and sent back and forth.

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

Claim from Brian Armstrong

Report · Money

Coinbase Business lets companies accept stablecoin payments from humans and AI agents

Coinbase has a product called Coinbase Business that helps companies accept stablecoin payments out of the box, from both humans and AI agents.

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

Claim from Brian Armstrong

Recommendation · Money

Choose use cases carefully for first optical compute demos

Ilker's take-home message: for the first demonstrations of useful optical compute, carefully choose the use case and algorithm to run on optical missions so that a tangible advantage compared to electronics can be created. He notes that using general purpose electronic hardware for diffusion-based image generation creates big redundancy because a single trained neural network is applied repeatedly, sometimes up to a thousand times, to create a single sampling of a single image.

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

Recommendation from Ilker (guest)

guest's recommendation from experience

Anecdote · Money

Armstrong now tags agents in Slack and sends PRs instead of pinging the team

Instead of pinging the team in Slack with requests, Armstrong now tags an agent in Slack with the comment or drops it into the internal harness and sends the team a PR for review. He finds working this way addictive because ideas and bugs get done rather than queued for the team — and AI tooling now even allows the CEO to get back in and ship features.

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

Claim from Brian Armstrong

Anecdote · Money

Armstrong planned a complex feature with a high-end model, then ran 10 parallel agents on phase one

Armstrong describes asking a high-end model to plan a relatively complex feature broken into three phases of 10 pieces each, then instructing it to spin up 10 separate agents to execute phase one's pieces in parallel and recommend cheaper agent models where appropriate. The parallel agents used open source models and Grok, and completed all 10 phase-one pieces shortly before the podcast.

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

Claim from Brian Armstrong

Anecdote · Money

Armstrong's parallel agent workflow and the addictive shift to sending finished PRs

Armstrong describes asking a high-end model to plan a complex feature broken into three phases of 10 pieces each, then executing phase one by spinning up 10 separate agents in parallel. He finds working with AI agents this way addictive, describing the shift from pinging the team to sending finished PRs as a magical moment. The host adds that non-technical people who adopt AI agents best experience it like a video game, leveling up their capability set through recursive loops.

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

Claim from Brian Armstrong

Recommendation · Money

Optical computing strategy: application-specific ASICs, co-designed algorithms, and minimal digital interfaces

The first approach for optical hardware should not be general-purpose computing hardware like CPUs, but ASIC-type application-specific approaches where bit depth, precision, and data rate can be minimized. The opportunity in optical computing is designing hardware and algorithms together; the team hand-tuned their generation algorithm and repeatedly reused the same weights to tap into where optics is useful. The usefulness of optics depends on how small the interface with electronics/digital environment is; if information can be looped directly in the optical/analog system for longer time steps, the advantage could scale from the demonstrated ~7x to hundreds or more. When a large number of parameters or a large amount of data compared to the number of parameters must be transmitted to the

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

Recommendation from Ilker

The ~7x to hundreds scaling estimate is the speaker's speculation; the advantage threshold claim is their heuristic.

Principle · Money

Use what is native to the computing form factor

Alok suggests that when taking pieces of brain-inspired design onto other computing substrates, whatever is native to that form factor is probably the right way to improve it.

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

Claim from Alok (host)

host's heuristic, offered with hedging ('probably')

Principle · Money

You want some of the noise, not all of it

The hard part of using noise in computation is that you want some of the noise, not all of it, because system noise can muddy the picture.

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

Claim from Alok

Principle · Money

Co-design models with the physics of new hardware

The speakers argue you should build models around the specific physical behavior of new hardware systems to co-design something with an interesting result, working within the system's strengths and weaknesses rather than blindly applying an architecture. When starting from a unique model architecture instead, the advice is to pick the right physical manifestation to bring it to life. Whether to iterate model-first or hardware-first depends on where your angle comes from.

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

Claim from Y Combinator Paper Club

presented as endorsed heuristics

Principle · Money

AI eliminates tasks, not people

Armstrong argues AI increases productivity and eliminates tasks rather than people: existing employees get much more done and the pace of everything accelerates. He does not believe companies will shrink headcount to 10 people; instead AI increases the velocity of the existing team.

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

Claim from Brian Armstrong