Product Hunt Picks · Nov 8: The Model That Sees and Fixes Its Own Work

So today's number one on Product Hunt is kind of interesting. A hundred votes, not a lot, but the name really grabs you. Inkling 2 Small — they say it can see what it builds, and fix it. Li Bo, what's the deal here? Let me start with the bottom line: it's a model from Thinking Machines Lab, Mira Murati's outfit. 594B total parameters, MoE, only 17B active. But the headline isn't the parameters — it's built for long-horizon coding and visual tasks. It can see what it builds — how should I picture that? It generates a page, then looks at the result itself? Exactly. It takes the rendered output as input and judges for itself what's wrong. Traditional models write it and hand it in. This one writes it, then grades its own homework first. So that solves the frontend pain of tweaking until you lose your mind. Normally after it generates, you still have to take screenshots and compare by hand. The key is the loop. Before, when an agent wrote code, the visual feedback had to be pasted back in by a human. Now the model looks at the image itself and iterates on its own. That's what makes long-running tasks actually work. But 17B active — that can't be cheap to run, right? Can small teams even use it? It runs on Tinker — inference and fine-tuning bundled together. Small teams don't have to train it themselves, they just call it. That's the clever part: the model and the platform are sold as a package. Okay, I get why it charted — that image of 'it looks at its own work and fixes it' is just so vivid. People want to see an agent finish the job by itself. Number two, Sage2. From Levanto Labs — it turns real-time conversation directly into structured data and actions. I really like this one. It doesn't sell a UI, it sells an API. Developers can plug it in and parse dates, amounts, durations from a conversation — and track how intent shifts. Isn't that just information extraction? That's been around for ages. Not the same. The old generation transcribed the whole thing first, then extracted. This one is streaming. Mid-conversation, if the intent changes, it knows on the spot what the next step should be. So the use case is like customer service? The user changes their mind halfway through? Bookings, orders, expense reports — all of it. The key is it treats 'in-conversation state' as a first-class citizen. That design philosophy is worth more than just pulling out fields. 99 votes, second place — just one vote behind number one. Developers really do vote with their wallets for this kind of infrastructure tool. And number three — this isn't a tool, it's a linguistics experiment. Deduplication. Asking which meaning has taken over a word. Made by HackerNoon, and they're really rigorous about it. They grabbed 719 words that the tech world has 'stolen,' across 14 domains, and used Wikipedia page views to figure out who wins. Bug used to mean an insect, mouse used to mean a rodent, token used to mean a coin. I know this well — I get tripped up by these words every day. It scores each pair: does the original meaning still hold, is it a tie, or has the new meaning completely taken over. The data's actually fascinating — you can see how fast technical vocabulary erodes things. Free and open, plus a word cloud and leaderboard. It's not a product, it's cultural observation — but 93 votes says people love it. Think about it — something this small charts because it hits a collective nerve. Our whole industry is constantly changing what words mean. So to sum up — what do today's three have in common? Two are AI going deeper, one is looking back at language. But the common thread is they all deal with 'the gap between meaning and reality.' The model has to see reality, the API has to catch intent, the word list has to see how meaning shifts. So today isn't a pile of tools, it's a pile of understanding layers. Everyone's teaching machines to understand what's actually happening, more precisely. Right, and none of them build end-to-end products — they build capability layers. That's a clear shift on Product Hunt this year. So what's today's pick? If you had to choose just one. I'd go Sage2. Low barrier, you can hook it up and try it today. Inkling is strong, but you're tied to the platform — I'd wait and see. I agree — streaming intent parsing like Sage2 is something anyone building a conversational product should really check out. That's all for today.