←
Eugene Shvarts580 posts
Eugene Shvarts avatar

Eugene Shvarts

@mathfax

building the future of ai @typesafeai

◷ Joined November 2015▧ 29 Photos and videos
Eugene Shvarts avatar
Eugene Shvarts@mathfax·11h

Computer vision turned eyes that comprehend the world into a digital commodity, and opened up so many use cases that could never have been practical with a human performing each one. We're on the cusp of this revolution for real-world decision-making.

We worked with @typesafeai to test Jev in the field. Our Maintenance App lets apartment maintenance techs photograph appliance labels instead of typing in model and serial numbers. The problem: if a photo is unreadable, we need to catch it before the technician walks away. We tested Jev, Sol and Opus on 140 labeled photos to see how quickly and accurately they could flag photos that needed a retake.

Post attachmentPost attachment
Treg avatar
⟳ Eugene Shvarts retweeted
Treg@treg_ai·Nov 17

Introducing Jev @typesafeai powered jobs-api Try it at treg.to/jobs-api - One API to any jobs data - $0.03 to run a full job hunt - Real salary & Hiring manager data - Plugin to any agent for your dream job Fully open source, 0% markup Git Repo below 👇

1:01
Harry Stebbings avatar
⟳ Eugene Shvarts retweeted
Harry Stebbings@HarryStebbings·Nov 16

Why I am sick of the status games around researchers in San Francisco “I am so sick of the status games, especially in SF, where ML researchers are a higher tier than ML engineers, which are a higher tier than software engineers. I think that is completely messed up. My hope is that we can create a new type of role: a software researcher, someone who actually pushes the boundaries of software.” @CompleteSkeptic Love to hear your thoughts @vboykis @eugeneyan @sh_reya @jeremyphoward

0:20

So Diogo Almeida is one of the most fascinating CEOs I have ever interviewed. born in the Philippines 2008: wins bronze for the Philippines at the International Mathematical Olympiad 2013: starts his career as a software engineer at Amazon 2017: joins Google Brain, working on cutting-edge deep learning research 2018: leaves Google Brain and takes roughly 18 months off to figure out what makes him happy travels the world, experiments with improv, art and meditation becomes a top-30 League of Legends player discovers that none of it makes him particularly happy 2020: joins OpenAI when it's still a relatively obscure research laboratory helps develop reinforcement learning from human feedback (RLHF), the technique that transforms how people interact with AI 2022: co-authors the landmark InstructGPT paper the work becomes foundational to ChatGPT November 2022: ChatGPT launches and becomes one of the fastest-growing consumer products in history 2023: contributes to GPT-4, one of the most important model releases in AI history but starts questioning the fundamental assumptions behind the technology he helped build 2024: leaves OpenAI while ChatGPT is exploding in popularity people think he's crazy for walking away co-founds TypeSafe AI with Erik Gafni and Sasha Sheng spends two years in stealth developing an entirely new class of AI models September 15, 2026: launches Jev, TypeSafe's first System One model Company makes $100M in revenue in the first 7 days of launch raises a reported $40M seed round led by DCVC October 2026: a reported $870M Series A at a $7.5B valuation, backed by a16z argues that the next breakthrough in AI will come from making intelligence cheap, reliable and autonomous My notes from my discussion with @CompleteSkeptic below: 1. Why Jevons' Paradox Does Not Cut It. We Are Going to 10x That Diogo expects a "mega Jevons paradox", with usage that dwarfs the token consumption of LLMs and AI running in the background everywhere. He expects people to stop seeing AI as something that "slops crap at them" and start seeing it as something that makes life better. 2. Why SaaS Incumbents Are Better Positioned Than OpenAI and Anthropic in Many Application Spaces SaaS companies know their customers' workflows and have the scale to fund the upfront engineering that real automation requires. Diogo argues their edge comes from being closest to the problems, not from distribution. Teams inside products like TurboTax know where to improve the interface, while OpenAI and Anthropic are "still living in their ivory tower." 3. Why We Do Not Hire From OpenAI Diogo sees getting into OpenAI as an anti-signal and says its researchers love algorithms and scaling but hate data. Data is a low-status topic at most labs, and the few people focused on the right task would want to pursue their own. That leaves researchers obsessed with algorithms, which he likens to throwing money in the furnace. 4. How the AI Bubble Could Burst and Where the Demand Shock Could Be Part of today's AI demand is driven by FOMO, from people believing they are getting good results. Diogo finds many buyers miscalibrated and overestimating what AI can do. Once they face the realities of its capabilities, demand could fall quickly. 5. What Metric Did ChatGPT Optimize for in the Early Days? Before Claude Code, ChatGPT was optimized for monthly active users rather than the best chatbot experience, which is how sycophancy emerged. Companies are now optimizing for the MAU equivalent in coding agents. Without a dedicated model, Diogo doubts they will produce a comparable one soon. 6. Why the Talent Market Is Not Crazy for Us and We Don't Pay for Mercenaries TypeSafe pays people good amounts but not the highest, which filters out hired guns. Diogo says Anthropic is suffering from mercenaries because it pays more than anyone else, and an entire market now exists to prepare candidates for Anthropic culture interviews. 7. We Wanted Sonya Huang More Than We Wanted Sequoia Diogo did not want Sequoia's money and brought the firm in because of Sonya. He values her particular shape of brilliance, which offers a perspective no other advisor gives him. One TypeSafe value is to make great decisions slowly, and he sees her as a valuable thought partner for that. (links in comments)

Harry Stebbings avatar
⟳ Eugene Shvarts retweeted
Harry Stebbings@HarryStebbings·Nov 16

So Diogo Almeida is one of the most fascinating CEOs I have ever interviewed. born in the Philippines 2008: wins bronze for the Philippines at the International Mathematical Olympiad 2013: starts his career as a software engineer at Amazon 2017: joins Google Brain, working on cutting-edge deep learning research 2018: leaves Google Brain and takes roughly 18 months off to figure out what makes him happy travels the world, experiments with improv, art and meditation becomes a top-30 League of Legends player discovers that none of it makes him particularly happy 2020: joins OpenAI when it's still a relatively obscure research laboratory helps develop reinforcement learning from human feedback (RLHF), the technique that transforms how people interact with AI 2022: co-authors the landmark InstructGPT paper the work becomes foundational to ChatGPT November 2022: ChatGPT launches and becomes one of the fastest-growing consumer products in history 2023: contributes to GPT-4, one of the most important model releases in AI history but starts questioning the fundamental assumptions behind the technology he helped build 2024: leaves OpenAI while ChatGPT is exploding in popularity people think he's crazy for walking away co-founds TypeSafe AI with Erik Gafni and Sasha Sheng spends two years in stealth developing an entirely new class of AI models September 15, 2026: launches Jev, TypeSafe's first System One model Company makes $100M in revenue in the first 7 days of launch raises a reported $40M seed round led by DCVC October 2026: a reported $870M Series A at a $7.5B valuation, backed by a16z argues that the next breakthrough in AI will come from making intelligence cheap, reliable and autonomous My notes from my discussion with @CompleteSkeptic below: 1. Why Jevons' Paradox Does Not Cut It. We Are Going to 10x That Diogo expects a "mega Jevons paradox", with usage that dwarfs the token consumption of LLMs and AI running in the background everywhere. He expects people to stop seeing AI as something that "slops crap at them" and start seeing it as something that makes life better. 2. Why SaaS Incumbents Are Better Positioned Than OpenAI and Anthropic in Many Application Spaces SaaS companies know their customers' workflows and have the scale to fund the upfront engineering that real automation requires. Diogo argues their edge comes from being closest to the problems, not from distribution. Teams inside products like TurboTax know where to improve the interface, while OpenAI and Anthropic are "still living in their ivory tower." 3. Why We Do Not Hire From OpenAI Diogo sees getting into OpenAI as an anti-signal and says its researchers love algorithms and scaling but hate data. Data is a low-status topic at most labs, and the few people focused on the right task would want to pursue their own. That leaves researchers obsessed with algorithms, which he likens to throwing money in the furnace. 4. How the AI Bubble Could Burst and Where the Demand Shock Could Be Part of today's AI demand is driven by FOMO, from people believing they are getting good results. Diogo finds many buyers miscalibrated and overestimating what AI can do. Once they face the realities of its capabilities, demand could fall quickly. 5. What Metric Did ChatGPT Optimize for in the Early Days? Before Claude Code, ChatGPT was optimized for monthly active users rather than the best chatbot experience, which is how sycophancy emerged. Companies are now optimizing for the MAU equivalent in coding agents. Without a dedicated model, Diogo doubts they will produce a comparable one soon. 6. Why the Talent Market Is Not Crazy for Us and We Don't Pay for Mercenaries TypeSafe pays people good amounts but not the highest, which filters out hired guns. Diogo says Anthropic is suffering from mercenaries because it pays more than anyone else, and an entire market now exists to prepare candidates for Anthropic culture interviews. 7. We Wanted Sonya Huang More Than We Wanted Sequoia Diogo did not want Sequoia's money and brought the firm in because of Sonya. He values her particular shape of brilliance, which offers a perspective no other advisor gives him. One TypeSafe value is to make great decisions slowly, and he sees her as a valuable thought partner for that. (links in comments)

TypeSafe AI avatar
⟳ Eugene Shvarts retweeted
TypeSafe AI@typesafeai·Nov 16

how to Jev • you pay for the `state` part of your request once, no matter how many questions you ask about it • so it's usually cheaper to ask every question you might want answered, even before you know if you need the answer • then use code to sort it out take in all the information you might need, and filter down to what's important! reliable automation means covering all your bases

The answer has nuance! The way Jev charges for input tokens is: (state+(sum of all questions)) The important bit here is that you only get charged once for your state tokens no matter how many questions you run in parallel. Within a single call, two questions do cost more than one. But if you ask those two questions as separate requests with the same state, you pay more overall, because you pay for the state again in the second request. In cases where you are more cost-sensitive than speed-sensitive, it makes sense to do analysis on how often the follow-up questions are used in your use-case. If some follow-ups are needed very rarely, then you can save on cost by running them only when actually needed. But in cases where speed is more important than cost, then batching pretty much always wins (although there's some nuance there as well, such as a single super-huge follow-up question slowing down the call, so shunting it to its own call could make sense, depending on the frequency that that answer is needed!)

Diogo Almeida avatar
⟳ Eugene Shvarts retweeted
Diogo Almeida@CompleteSkeptic·Nov 14

surprisingly few people realize state is shared and asking a lot of Qs is very efficient

The answer has nuance! The way Jev charges for input tokens is: (state+(sum of all questions)) The important bit here is that you only get charged once for your state tokens no matter how many questions you run in parallel. Within a single call, two questions do cost more than one. But if you ask those two questions as separate requests with the same state, you pay more overall, because you pay for the state again in the second request. In cases where you are more cost-sensitive than speed-sensitive, it makes sense to do analysis on how often the follow-up questions are used in your use-case. If some follow-ups are needed very rarely, then you can save on cost by running them only when actually needed. But in cases where speed is more important than cost, then batching pretty much always wins (although there's some nuance there as well, such as a single super-huge follow-up question slowing down the call, so shunting it to its own call could make sense, depending on the frequency that that answer is needed!)

JohnPhamous is ooo avatar
⟳ Eugene Shvarts retweeted
JohnPhamous is ooo@JohnPhamous·Nov 13

using pointer velocities to predict which button the user is going to click on

0:11
Eugene Shvarts avatar
Eugene Shvarts@mathfax·Nov 13

One of the fun parts of teaching Jev is that the architecture allowed for extremely flexible pattern design, and the cheap intelligence optimized for always asking more then paring down (just like real intelligence!). For example, caching at the System One layer, maps to asking all your questions together, so you speculatively ask even ones you don’t know you’ll need yet! This helps flex the muscle of figuring out everything you need to ask in the first place.

One of the most important Jev patterns to learn is what we sometimes call "speculative questions". In the majority of cases, it's faster and even often cheaper to batch questions together in to a single call, *even when you don't yet know if you'll need the answers!* Why? Many reasons: Speed! Oftentimes you can accomplish the entire task in a single call, instead of multiple sequential calls in different logic paths. Cost! You only pay for tokens in state once per call, so batching questions also avoids paying for the state multiple times. When your state is large, batching questions often saves cost compared to sequential calls that duplicate state. Questions never affect the output of other questions in the same call, so accuracy will not suffer.

Post attachmentPost attachmentPost attachmentPost attachment
Eugene Shvarts avatar
Eugene Shvarts@mathfax·Nov 12

One of the dreams (pun intended) of smart software is that if you can describe it, you can build it! Why should detecting people at the edge of the video be any harder than just asking?

Introducing Wisp, open-source content-aware video player that softly fades the edges of a video into the background for a dreamlike look. It uses Jev by @typesafeai to detect people, animals and faces and keeps them sharp as they get close to the edges.

0:12
Eugene Shvarts avatar
Eugene Shvarts@mathfax·Nov 12

1,000 requests → down to 11 30+ seconds → about 1 🔥 Half a cent per search.

Eugene Shvarts avatar
Eugene Shvarts@mathfax·Nov 12

So put your eng hat on, because using Jev well is just programming. One Choice question can rank 100 listings at once, so we split the 1,000 into 10 groups, keep the top 5 from each, and one more call scores the 50 finalists.

Eugene Shvarts avatar
Eugene Shvarts@mathfax·Nov 12

The naive way is one Jev call per listing. Each one's fast, but 1,000 of them back to back blow past 30 seconds, and Google Sheets kills any formula that runs that long 💀 sounds like a skill issue.

Eugene Shvarts avatar
Eugene Shvarts@mathfax·Nov 12

Never read an Airbnb listing again. Tell Jev what you want in plain English, and it reads 1,000 Cape Town listings and hands you the 10 best, scored out of 10, in about a second. Right inside Google Sheets 🏠 (trying something new with our demos!)

Dan Robinson avatar
⟳ Eugene Shvarts retweeted
Dan Robinson@danlovesproofs·Nov 10

Nice unlock from jev: we can show you what parts of your codebase are relevant to a vuln scan and let you fine-tune the targeting. The classifications are fast enough and cheap enough that you can support "insert common sense here" functionality inside of a UX that needs to be snappy. This works for repos with 100k+ files.

Post attachment