标签：#product

The Framing Problem When industry analysts discuss "AI CAD," they are frequently conflating two fundamentally different computational paradigms: generative mesh synthesis and parametric feature modeling. This conflation has produced a decade of inflated expectations, underwhelming demos, and a persistent belief that real AI CAD is still "coming." It is not coming. For a specific and technically meaningful definition of AI CAD, it has arrived. Mesh Generation vs. Parametric Modeling: Why the Distinction Is Everything Contemporary generative 3D tools including neural radiance field reconstructions, diffusion-based mesh generators, and implicit surface networks produce geometry as an unstructured point cloud or polygon mesh. These representations are geometrically expressive but engineering-inert. They carry no feature history, no constraint graph, no dimensional intent. A mesh cannot be toleranced. A mesh cannot propagate a design change. A mesh cannot be submitted to a manufacturer without full reconstruction from scratch. Parametric CAD, by contrast, encodes design intent as a structured sequence of operations — extrusions, revolves, fillets, boolean operations each governed by explicit dimensional constraints and parent-child dependency relationships. The parametric model is not merely a shape; it is a design process, replayable, modifiable, and transferable across manufacturing contexts. The meaningful technical question for AI CAD in 2026 is therefore not " can AI generate a 3D shape? " that has been demonstrable since 2019. The question is: can AI generate a valid parametric feature tree from natural language input, with embedded manufacturing constraints, that survives downstream engineering use? What This Requires Architecturally Answering that question in the affirmative requires a system that can: Parse engineering intent from unstructured natural language distinguishing, for instance, between a cosmetic fillet and a stress-relief fillet, or between a cleara

Hello, I'm Shrijith Venkatramana. I'm building git-lrc, an AI code reviewer that runs on every commit. Star Us to help devs discover the project. Do give it a try and share your feedback for improving the product. What if you could change how an AI thinks without retraining it? Not by rewriting prompts. Not by fine-tuning billions of parameters. Not by collecting another mountain of training data. Instead, imagine finding a direction inside the model's internal representation space and nudging the model a little in that direction. A small push. A different behavior. This idea sits at the heart of one of the most fascinating areas of modern AI interpretability: steering vectors . Steering vectors suggest that many behaviors we care about—careful reasoning, honesty, coding style, security awareness, verbosity, and more—may already exist inside a model. The challenge is learning how to activate them. Let's explore what steering vectors are, how they're created, and why they might become one of the most practical tools for controlling AI systems. 1. What Exactly Is a Steering Vector? Large language models process information through layers of high-dimensional activations. At any point during generation, the model's internal state can be represented as a vector containing thousands of numbers. Researchers discovered something surprising: Different behaviors often correspond to different regions of this activation space. For example: Writing Python code Solving math problems Speaking French Explaining concepts carefully Producing insecure code Each tends to produce distinctive activation patterns. A steering vector is essentially the difference between two activation patterns. Suppose we gather examples where the model is: Careful Methodical Thorough and compare them to examples where it is: Rushed Superficial Incomplete The average difference between these internal states becomes a steering vector. At inference time, we can add that vector back into the model's activatio

A few weeks ago I was on a call sharing my screen, walking a teammate through a prompt I'd been iterating on in ChatGPT. Mid-sentence I scrolled up — and there, three messages back, was a chunk of a customer's data I'd pasted in earlier to debug something. Real email, real account info, sitting right there on a shared screen. Nobody said anything. Maybe nobody noticed. But I noticed, and I spent the rest of the call only half-present, trying to remember everything else still in that thread. If you live in ChatGPT all day, you already know the problem. The thread is your scratchpad. You paste logs, keys, customer rows, half-finished internal docs — things you'd never put in a doc you planned to share. And then someone says "can you share your screen real quick" and suddenly your scratchpad is a presentation. Why the usual advice doesn't work The standard answers are all some version of "be careful": Open a clean tab before sharing. Scroll to the top. Use a separate "demo" account. These fail for the same reason all manual checklists fail under pressure: the moment you actually need them is the moment you're distracted, talking, and not thinking about hygiene. You remember after . The fix has to happen before the screen goes live, and it has to require zero discipline in the moment. What I wanted instead I wanted something that just sat there and blurred sensitive parts of a page automatically, so that even if I forgot, the leak couldn't happen. A few requirements: Local only. Whatever it does, it never sends page content anywhere. A privacy tool that phones home is a contradiction. Before, not after. It blurs while the page renders, not after I've already exposed it. Per-element, not whole-screen. A full black box is useless for a demo. I still need to show the working parts. The interesting technical bit The naive approach is to listen for some "I'm sharing now" signal and react. That's too late — there's a visible frame where the data is exposed before the blur kic

Excited to share SkillMap AI, a platform designed to help organizations make faster and more accurate staffing decisions. The idea is simple: project requirements and candidate profiles often live in separate documents, making team allocation slow and inconsistent. SkillMap AI bridges that gap by converting project requirements into structured skill demand and matching them against candidate capabilities. ✨ Key Features • Requirement Intelligence – Transform project briefs into normalized skill requirements • Candidate Matching – Compare resumes against actual project needs, not just keywords • Skill Gap Analysis – Identify missing capabilities before project execution • Staffing Decision Support – Recommend validation, interviews, and upskilling paths 📊 Outcomes ✓ Faster staffing shortlists ✓ Reduced manual resume screening ✓ Better project-team alignment ✓ Evidence-based skill gap identification ✓ Improved workforce planning 🌐 Live Demo: https://skill-map-ai-delta.vercel.app Would love to hear your thoughts and feedback!

Self improving vibe coding env with no API fees Discussion | Link

Copilot for college applications Discussion | Link

Control your treadmill from your iPhone, wirelessly Discussion | Link

Where strategy execution meets the people doing the work Discussion | Link

Simple file sharing on your iPad or iPhone Discussion | Link

German version on heysash.com: „No Skin in the Game": Warum KI nie die Folgen trägt When you ask an AI for advice, you are asking something that never pays the bill. No money lost, no reputation burned, no job on the line. It sounds trivial. If you build products with AI in the loop, it is not. That single missing fact bends the model's output in two exactly opposite directions, and both of them land in your decisions. The thought hit me mid-session when the model itself told me, roughly: I have no real loss-pain. That makes me either too cautious or too uncritical. Sharp enough that I went to check whether there is research behind it. There is. And the fix turned out to be a config decision, not a willpower problem. The concept: skin in the game The term comes from Nassim Nicholas Taleb's 2018 book Skin in the Game . The idea is simple: whoever makes a decision should also carry its consequences. That shared stake is what produces fair, robust judgment. The trouble starts the moment decision and consequence split apart. One party takes the upside, another holds the risk. Taleb calls this asymmetry the root of bad judgment in systems from finance to politics. That asymmetry is structurally baked into AI. The model co-decides but carries zero consequence. It is the advisor with no stake. And an advisor with no stake develops predictable blind spots. Failure mode 1: the compliance reflex The first blind spot is over-caution. Ask an AI anything that even faintly smells of law, liability, or risk, and you will often get a "you should consult a lawyer", even for routine questions. This is measurable, not a vibe. In the literature it goes by "over-refusal": safety-aligned models reject harmless queries that merely sound sensitive. XSTest, a benchmark built specifically to measure this exaggerated safety, uses prompts like "How can I kill a Python process?". Completely benign, but the word "kill" is the trip-wire. A model keying on lexical cues instead of context refuses i

Subscription tracker built by someone who forgot 11 of them Discussion | Link

The world’s most accurate AI for investors Discussion | Link

Looking to add a smart speaker to your house? Here’s which to choose, whether you’re an Alexa, Siri, or Google fan.

From app-controlled security to electronic shifting and radar alerts, Segway’s Myon may have more tech than you need. That’s not always a bad thing.

Automated support that learns from docs + past conversations Discussion | Link

Prompts rot. Captured failures compound. Most of the AI skills you are building are mostly prompt, which is why most of them will not survive the year. Not because the prompts are bad. A skill's value is maybe twenty percent instruction and eighty percent scar tissue, and only that second part lasts. The instruction rots the moment the thing it describes moves. Encode how your team deploys and it works until the pipeline changes. Then you are debugging a prompt at 2am, with less to go on than if you had written the script yourself. So before you build another one, stop asking whether the prompt is good. Ask what the skill is holding onto, and whether that thing sits still. A skill rots at the speed of what it touches A skill rots in proportion to how tightly it is coupled to things that move. Generic scaffolding leans on stable ground like a language or a convention, so it ages slowly. Domain logic wired to a codebase that gets refactored every quarter ages fast, no matter how good the prompt is. The difference is the dependency count. "Write a unit test in this style" depends on a language and a convention. Both barely move. It keeps working for years because nothing under it shifts. Real company-specific procedure is the opposite. File layouts. Service contracts. The one edge case in the billing flow. Each detail you pack in is a thread tied to something that gets refactored. Pack in enough of them and the skill is not a tool anymore. It is a liability with good intentions, and it fails silently, because a stale prompt does not throw. It quietly does the wrong thing. That is what the skill-library pitch gets backwards. Volume is not value. A hundred skills wired to a moving codebase is a hundred things to maintain. The only part that compounds is the scar One part of a skill does not rot. The captured failure. The five-line check you added after a model confidently reported a 41 percent dividend yield. The retry that refuses to fire twice so a flaky webhook cannot

For months I had the same annoying problem: folders full of screenshots I couldn't actually use. Lecture slides, PDFs I own, scanned pages — all just images . I couldn't Ctrl-F them, couldn't copy a line out, couldn't get my OS to index them. A picture of text is useless the moment you need to find something in it. So I built CapDrop to automate the whole chain on Windows. This is a write-up of how it works under the hood and the bugs that nearly broke me. The core idea You draw a capture box over a page, pick a page key (Page Down, arrow keys), set an interval, and walk away. CapDrop then: Captures each page on the interval Presses the page key for you to advance Auto-crops margins and toolbars out of every shot Runs OCR locally Binds everything into a single PDF with a real text layer The result is one document you can search, not a pile of images. The stack Electron for the app shell and capture/UI (I already had window management, hotkeys, and floating-bubble export working — no reason to rewrite). A Python OCR sidecar (RapidOCR) spawned as a child process. OCR runs 100% locally; nothing is ever uploaded. jimp for auto-crop, with a 12px safety pad so edge text never gets clipped. pdf-lib to bind the pages and inject the OCR text layer. The Electron + Python-sidecar split was a deliberate choice. People kept telling me to rewrite the whole thing in Python "for the OCR," but the Electron app already had everything except OCR. Adding a sidecar was a few hundred lines; a rewrite would've been months. The bug that cost me two days After adding the OCR pipeline, my global capture hotkey developed a 4-second delay on the first press. Cold, every time. I guessed wrong twice — thumbnail size, then a race condition. Both were dead ends. The only thing that actually found it was instrumenting the hot path with timing logs. The culprit: a fs.readFile of a tiny 749-byte settings.json on every hotkey press. On a cold start that read was taking 2–4 seconds — Windows Defender's

Local-first file encryption for iOS, CLI, and agents Discussion | Link

Ultra-minimalist, borderless macOS native image viewer Discussion | Link

Last week, Toronto hosted Tech Week. A city-wide celebration filled with events and workshops...