Shipping code shouldn’t mean hours of manual testing. @Simulithic (YC F26) sends AI agents through your app like real users, across browsers, devices, and account types. It tells you what broke without you writing a single test script. The teams that ship fastest will be the ones that can verify as fast as they build. Get started today: simulithic.com

I reverse engineered Pocket and made it 10X better. Right now: the app is full of ads, the microphone is low quality, it constantly misspells peoples names, and has no agentic features. I changed that.

We all consider Anthropic an AI company. But what if it's a cloud?

Doing a hackathon with @eduardfausgil & @ayushrgarg was the best decision I’ve ever made I was in India, about to give up on tech back in 2022, met these 2, learned how to build cool things with them, and it changed my life (Picture below is approximately 4.5 years apart lol)

value accrual at the model layer is splitting in two. for frontier discovery, it concentrates in a handful of labs. for commoditized workflows, it leaks out to whoever is cheapest or closest to the customer. the reason is that the value of being one model generation ahead varies enormously by task. for problems with known answers, intelligence eventually becomes abundant. for problems without known or verifiable answers yet, the strongest model can be worth orders of magnitude more than the next best. frontier discovery is novel research and long-horizon agentic work where nobody knows the answer yet. bio appears to be a large one over the next few years for ex. models are becoming exceptionally strong at prediction, and the value of a single hit dwarfs the cost of the tokens that found it. expect the labs to start running their own experiments, with frontier models choosing what to test. the feedback loop here isn’t just better models producing better predictions. it’s better predictions determining which experiments get run, generating new ground truth that makes the next round of predictions better. openai and anthropic are best positioned here, with google close behind. buyers of discovery pay for the strongest model, so being a generation behind costs far more than it does anywhere else. the lead compounds given the labs use their best models to accelerate building the next ones, they have the compute to do it at scale, and they keep those models internal long before exposing an api. commoditized workflows are the long tail of tasks a model one generation old already handles well. quality saturates and cost decides. deepseek and other chinese labs tend to own the cost-quality pareto frontier, partly out of necessity given their relatively compute-constrained conditions, meaning efficiency has become integral to the research agenda. the western frontier labs have their own edge in distillation, as they can leverage the strongest teachers at the logit level, while everyone else distills from a weaker teacher or from sampled text. but they only win this tier if they choose to compete on price. even the pareto leader doesn’t capture much per task here, because the weights are generally open. anyone can self-host them or finetune them on their own data, and as a result the model layer’s share shrinks to an inference margin. but thin margins on an enormous volume of workflows can still add up to substantial value capture. the model layer becomes a scale business rather than a pricing power business. cost isn’t the only place value moves. whoever owns the customer relationship also owns the workflow, the context, and increasingly the feedback that determines what good performance actually looks like. a caveat then is taste. general taste lives in the weights, but most “commodity” workflows are full of judgment calls specific to one business. the ground truth for those comes from outcomes the labs never see. if you’re building on top of the labs, don’t compete on general intelligence. own the loop that turns your customers’ outcomes into a reward signal.

I always knew I was going to do my startup at some point But it was the layoff from my internship in Canada a year ago (before Bree), that actually paved the path for it Planning to make more such videos in future :)

"Our initial evaluation suggested that GPT-6-Astra's score was in the late eighties followed closely by Opus 5.5. But, as we started to investigate the mistakes that the models might have made, the models, especially Astra, insisted that they were right and that it was the expert-supplied answers that were erroneous. When we went back to the experts it turned out, in multiple instances, that that the experts found some subtle errors buried deep inside their calculations. As they revised their answers, Astra's score started to inch up and it finally ended up with 98% (49/50)."

Reminder

Already showing people what they can build with Codex? Applications to become a Codex Ambassador are open again. We support you in hosting meetups or sharing what you’ve learned: developers.openai.com/commun…