@svpinoOct 23

Jev is really useful, actually. I have a script that receives a notification every time I get a new email. I use Jev to decide whether I care about that email or not (some sort of spam filtering). This is "personalized spam" detection. A layer on top of my email provider's spam filter. My goal is to catch emails that aren't necessarily spam but that I still don't care about. I classify new emails with Jev to decide whether they're spam. If Jev's confidence is above 90%, I delete the email. I've been running this for about 3-4 weeks (pretty much since Jev came out), and so far, I've deleted hundreds of messages. During the first week or so, I manually checked most of the emails the script discarded to make sure I wasn't missing anything important. I wasn't. The script is good.

@aakashguptaOct 22

71 companies now train their own AI models. Here's every one of them, on a single map. Ask most people who trains AI models, and they'll name three labs. The real list runs from Midjourney to a company that designs drugs. And xAI isn't even called xAI anymore. Here's what stood out: 1. The frontier changed names xAI is now SpaceXAI, and SpaceX bought Cursor. Meta's flagship is now Muse Spark, and it's closed. Google's Gemini 4 Argon sits at #1 on LMArena's text leaderboard (on a preliminary score). 2. China runs the busiest open models Last week, DeepSeek models handled 21.8% of all requests on OpenRouter, more than any other lab. Qwen, Kimi, GLM and MiniMax all shipped open weights this year. 3. Deals redrew the map AMD agreed to buy World Labs. Cohere and Aleph Alpha agreed to merge. OpenAI retired the Sora app. 4. A new category showed up Decision models never write text. They answer a typed question with a probability. TypeSafe's Jev is one, and the map already has four companies in the category, including Liquid AI and Together AI. 5. Specialists are everywhere Music, robotics, biology and world models all have their own model companies now, from Suno to Physical Intelligence to Chai Discovery to World Labs. 6. The long tail is huge Hugging Face now lists 3.1M models. The 71 companies on this map are the ones training their own. 7. The US still leads on volume Stanford's AI Index counted 59 notable models from the US in 2025, and 35 from China. Every company on this map needs people who decide what its model should get better at, how it should behave, and when it's ready to ship. That's the model PM job. I counted the openings: 46 model PM postings are live today at 17 companies, at a median posted base of $345K. None of them lists hands-on ML research as a must-have. I put the map, every opening and the roadmap in one guide. Which company would you add?

@omarsar0Oct 22

Recommended reading. I've been testing decision models such as Jev and Drex Max for research tasks (e.g., paper classification). Just published our initial exploratory work on it. Drex Max is a strong fit for our use cases, and we are already planning production efforts around it. Multimodal decision models also open up a whole range of exciting areas. We'll share more results and resources soon.

@omarsar0Oct 19

Decision models are going multimodal. Drex Max looks really strong. It can make fast decisions straight from screenshots, screencasts, scanned documents, scientific charts, and photos. This model is a really good fit for computer-use and browser-use agents, where every step starts with reading the screen. Nace reports 88.3% on Mind2Web (outperforms Jev 1.14). Drex Max also scores 67.6 on the Decision Index (Jev 1.14: 60.70).

@hammer_mtOct 16

My first 1m impression month! What worked you ask? Well... 1. adding hoboken, NJ to a map of best places in NY to live (they hate that) 2. hyping up @kieranklaassen's jev work (support your colleagues) 3. criticizing anthropic's cloud move (people hate change) 4. making fun of accenture's token budgets (lol) 5. telling people tokens will cost more than salaries in the future (true but people don't want to admit it) 6. teaching people about evals (be helpful)

@openrouterOct 15

Today, we are launching OpenRouter Armor in beta. Tool calls allow agents to interact with the world, and regex-based filtering often isn't enough to detect unwanted tool calls. Armor uses @typesafeai's Jev model to guard against out of policy tool calls before your agent executes them, with better performance than Claude Code's auto classifier at a fraction of the cost.

@harrystebbingsOct 15

For the last 5 decades straight, we have created (as a society) circa 20 decacorns per decade. In the last 12 months, we have created 42. The world is changing faster than ever before and if you do not stay up to date. You will be left behind. That is why I created this show with Jason and Rory. This is the only show in tech, you have to listen to every week. No egos, no politics, just tech. On the agenda: - Valar Atomics Sues Day One Ventures - Meta Returns to the AI Frontier With Muse Wonder - NVIDIA Eyes $30B Reflection Deal - Jev Raises $870M at a $7.5B Valuation My notes below with @jasonlk and @rodriscoll 1. We Generated 20 $10B Companies per Decade for the Last 5 Decades and Now 42 in a Year Tech produced roughly 20 decacorns per decade for five decades, yet 42 have been created in less than a year of this decade. Only about 130 tech companies have ever been worth $10B or more. Rising market caps explain part of the jump, but the pace is stunning. 2. Who Is Right and Who Is Wrong: Day One or Valar? X wrongly claimed Day One was suing its portfolio company, when Valar Atomics wants a declaratory judgment to move on. Waiving pro rata for Sequoia is common, but letting an existing investor in as a strategic is playing favorites. Still, once founders decide, early investors must get with the program because it is not their company. 3. Why Meta's New Model Is a Zero Meta's new model scores well on evals, but there is no API and it can only be used inside Muse. A model claiming parity at the same price, not cheaper and not usable by developers, is the height of performative evals. 4. Why We Should Lead Jev's Round at 8 Billion Jev is a super fast, super cheap classifier, accurate enough for maybe 10% of AI use cases. Cloudflare and Amazon can match it today, but the bet is that a focused team out of OpenAI pulls ahead while those efforts stay niche. If Jev becomes hyper accurate, it changes the game. 5. What Is the Single Greatest Potential for AI? Drug discovery is probably AI's largest opportunity, with drugs that dramatically extend longevity and quality of life. If AI companies can accelerate discovery and own the IP while others do the hard work, AI could reboot the industry. Trials would not speed up, but the prize is a GLP-1 out five years faster. 6. Haiku 5.5 Is the First Time AI Has Been Deflationary Falling token prices have been offset by consuming far more tokens, so AI costs have not fallen. Haiku 5.5 is almost as good as Sonnet for many use cases at a tenth of the price, and Sonnet 5.5 is twice as fast. After a 90% price cut, usage must rise 10 times for revenue to hold. 7. Why the SaaSpocalypse Is Worse Than We Thought The SaaSpocalypse passed only for companies well above $1B in revenue, and everyone without massive scale is living a worse one than in March or April. Leaders will acquire some companies and grind out the rest. A safe software IPO may now need $1.5B to $2B of ARR rather than $500M. (links in comments)

@typesafeaiOct 15

Known affiliation

we announced Jev 1.14 on 10.14 at 1:14pm PT 🔥 a new and improved model, along with experimental features for our dreamers: • images • a new box type • longer context surprise, Jev was already able to see you! 👁️ BUT our bar for production-grade intelligence is quite high, so builders beware, these are preview features. enter the ~*~ danger zone ~*~ at your own risk. 🚨

@wesbosOct 14

Jev images is here One question about an image that is 850x850 cost one hundredth of a cent - $0.000096936 about 10,000 images for a dollar. referencing the image more than once duplicates the token usage - so asking two questions in a single request is equivalent to uploading two images

@completeskepticOct 14

Known affiliation

Happy JevDay! But before announcements we have to talk philosophy… The needs of early adopters / tinkerers / dreamers are quite different from what is needed for extremely stable infrastructure. For the Jev release, we erred on the latter, but this release is for the former - with a gate so people can opt-in to the danger zone!

@kevinroseOct 14

so-so-so proud of @JustinMezzell @addison @maubrowncow for rebooting @digg. thousands of articles per day, tens of thousands of jev decisions, real-time clustering, new topics, digging is back, it's really a fantastic news resource, check it out if you haven't lately. cc: @alexisohanian

@deedydasOct 14

Modal is one of the only companies in history to grow 10x in <1 year on a 9 figure run rate base to $1B! It’s a privelege to continue partnering with them: it was the #1 investment for me when I joined venture and has become a key part of the way most companies scale, including Jev. Erik and Akshat are also some of the nicest, most thoughtful founders and now join the very exclusive club of IOI gold medalists x decacorn founders.

@completeskepticOct 14

Known affiliation

was at a community event and was asked a somewhat sad question: should developers be afraid of the labs eating the new possibilities jev opens up? with AI (system one) + software (system two): 1. there's no going back to a world without system one models 2. you control the software! 3. there's no way AI could beat the efficiency of this combo 4. you don't even send the model providers your whole program logic I think the days of "wrapper" apps will be seen as some of the darkest in software because people essentially did all of their programming in a system message and sent it over to a third party who was incentivized to bake that behavior into their 1st party product... keeping your logic in code and only asking AI for the minimum reverses that and keeps all of your logic yours. if you decompose enough, model providers can't even figure out what you're working on

@exm7777Oct 12

this is Jev but on STEROIDS and nobody is ready for it... output tokens are basically FREE now lemme explain to you what it means... you only pay for what you send it, $0.05 per million tokens, and the answer costs you nothing and it doesn't just read a little bit... it reads up to 384K tokens in ONE go contracts, logs, tickets, screenshots, charts, scanned pages, all at once, no chopping anything into pieces then it makes a decision in under a second - which support ticket goes where - which security issue gets flagged - which tool the agent picks - what the legal or finance doc actually says it leads on 13 of 16 enterprise decision benchmarks your agents used to guess because giving them all the context was too expensive... that excuse is gone now

@omarsar0Oct 12

Pay attention to decision models. From my recent tests, decision models can be extremely transformative for businesses. Many calls in an agent workflow are structured decisions, so paying frontier prices doesn't make sense. SubQ Snap from @subquadratic is trained with RL for those calls. It holds 384K tokens of context and leads Jev on 13 of 16 enterprise benchmarks, so it can decide with a whole repo or a stack of contracts in view.

@itsolelehmannOct 12

damn Jev just got lapped. Snap now has 12x Jev's decision context. and these decision models mog LLMs for micro-decisions where your agent needs an instant answer: > which ticket needs a human? > is this lead worth chasing? > does this contract clause break our rules? otherwise, you're just wasting time and tokens waiting for Claude to generate an answer when your agent just needs: > yes/no > a choice > or a score its basically like hiring a rocket scientist to sort support tickets. and with a massive 384K tokens, Snap can factor in more customer history, past deals and company rules when making those calls.

@svpinoOct 12

Decision-making models are the new hot thing in the industry. SubQ Snap is the latest model to come out, and it looks pretty strong: • Up to 384K tokens of context in a single call • Processes 100K tokens and returns a decision in <1 second • Uses sparse attention to skip unnecessary token comparisons • Handles text and images in the same context • Trained to handle structured, data-heavy decisions • Leads Jev on 13 of 16 Decision Index benchmarks

@aakashguptaOct 10

The guy who taught ChatGPT how to talk just raised $870 million to build AI that never speaks. Three weeks. That's how long his company has existed in public. Diogo Almeida spent four years at OpenAI. He co-created RLHF, the training method that turned a raw model into something that sounds like a person. Then he left, and on September 15 he came out of stealth with a $40M seed at a $200M valuation. On Thursday a16z priced his Series A at $7.5 billion. Same company, 24 days apart, at 37x the price. Here's the bet. His model Jev doesn't write anything. You feed it language and it hands back one decision from a list of options, with a probability attached. A fraud check doesn't need an essay. Software can't read prose anyway, which is why almost no AI actually runs inside production systems today. It all sits in chat windows talking to humans. TypeSafe's own math puts one decision at about $3 versus roughly $372 for the same check on a frontier LLM. At $3 you stop rationing AI calls. You put one behind every branch in your code. That's how 29.4% of the Fortune 500 showed up in three weeks with zero marketing. The man spent four years making machines sound human. His $7.5 billion bet is that the money is in machines you never hear from at all.

@aakashguptaOct 10

Microsoft just shipped an AI that makes decisions faster than you can blink. A blink takes about 100 milliseconds. This one takes 85. And Satya didn't pick this day by accident. A few hours before this post went up, a startup called TypeSafe closed an $870 million Series A at a $7.5 billion valuation. TypeSafe has existed publicly for 24 days. Their seed round in mid-September valued them at $200 million. 37x in under a month, for inventing the exact category Microsoft is now charting itself on top of. Quick explainer, since most people haven't met one of these models yet. A decision model doesn't chat. You give it a bounded question, "is this transaction fraud," "which team owns this bug," and it hands back an answer plus a probability in a fraction of a second. Software acts on it directly. Nobody reads a paragraph first. This category did not exist on September 14. TypeSafe launched Jev on the 15th. By October 1, OpenAI had a Decisions API out, Cloudflare had shipped Clef with free downloadable weights and a 39 millisecond response time, AWS and Perplexity had their own versions live. Now Microsoft wants the accuracy lead. Five of the biggest names in tech chasing a one-month-old startup, which should tell you how much money they think is sitting in this layer. Because the money part is wild. Jev charges $0.042 per million input tokens and output costs nothing. Meanwhile a frontier LLM grinds for 3 full seconds on tasks this new Microsoft model clears in 85 milliseconds. When a decision costs basically nothing and takes basically no time, you stop rationing decisions. Every support ticket, every log line gets its own ruling as it happens. The column worth staring at on Satya's chart is calibration. Two new hires. One is right 83% of the time and tells you exactly when she's unsure. The other is right 95% of the time and delivers every answer with total confidence. The first one you can hand real responsibility. The second one you babysit forever, because you never know which answers are the rotten ones. Microsoft-Decision-1 scores 92.2 on calibration, where 100 means the model's confidence perfectly matches how often it's right. That's the score that decides whether a human still has to sit next to the machine.

@omarsar0Oct 9

NEW: Microsoft also releases its System One model, Microsoft-Decision-1. What a crazy effect Jev has had in the space. Also, a very cool application of decision-making models to power LLM judges and improve scientific discovery pipelines (e.g., screening candidate hypotheses). I'll add this model to my ever-growing list of eval runs. In my ongoing evals, consistency and more complex decision-making are two areas where I see these decision models struggle. I'm also exploring a bunch of science applications at @dair_ai and will share when ready.