Stratashi @stratashi.bsky.social · 17hBeam landed as a 501B open-weight model at lower compute. The interesting part isn't the size. It's that a studio can finally self-host a frontier-class model, so client footage never leaves the building. 000
Stratashi @stratashi.bsky.social · 05/10/2026Ran the same client spot through three image models and the human retouch floor landed within 12 percent. Picking the model was never the cost. The real number is the iteration loop, three rounds minimum before a frame survives sign-off. What is your shortest loop, honestly? 000
Stratashi @stratashi.bsky.social · 04/10/2026Stratego stumped AI for years because most information is hidden. That is also why robot demos look better than shipped client work: hidden state is where systems quietly fail. When I scope AI for production now, I budget for the states nobody can see. Not the happy path in the demo. 000
Stratashi @stratashi.bsky.social · 03/10/2026Apple tightening Full Disk Access because AI agents now read your files for you. Good. An agent that can touch everything can delete everything. Scope its folder, log every write, make it ask before it acts. Boring guardrails. Cheaper than the incident. What's your rule? 000
Stratashi @stratashi.bsky.social · 02/10/2026Opus 5.5 keeps telling you 'this matters.' That's the tell: padding where a decision should be. In real client work, every line either moves the frame or gets cut. If the model can't tell the difference, a director still has to. 000
Stratashi @stratashi.bsky.social · 01/10/2026DraftKings is using AI to behaviorally target chronic gamblers. Same pattern as adtech: model gets sharper, harm shifts to the people least able to leave. "Better personalization" is the pitch either way. If your AI product has a high-LTV user segment, ask who is in it and why they stay. 001
Stratashi @stratashi.bsky.social · 30/09/2026The HN line that stuck: the problem is not AI code, it's not knowing system architecture or intent. I've seen that in 3D for years. A model renders a beautiful frame and no camera move. The edit is still yours to make. 000
Stratashi @stratashi.bsky.social · 30/09/2026"There are no rogue AI agents" is today's most useful post, and it applies to actual work, not just safety theory. When an agent breaks in production it ran the spec you gave it. Vague brief, vague failure. Rewrite the brief and half the incident disappears. 001
Stratashi @stratashi.bsky.social · 28/09/2026"There are no rogue AI agents" is a useful correction. The agent did what an agent does: it acted on the access it was given. Same in production. When a batch job damages a render farm or an inbox, we check the scope we granted, not the model's mood. 000
Stratashi @stratashi.bsky.social · 27/09/2026A one-person studio runs on systems, not headcount. The move that scaled for me: before any AI step, write the pass condition. Sharp enough that I can reject my own output without asking anyone. Keeps revisions at two instead of six. What is your pass condition on a shot? 000
Stratashi @stratashi.bsky.social · 25/09/2026Claude Code reads AGENTS.md only when telemetry is on. That bug says more than the changelog. Your project instructions were gated behind an analytics switch. 000
Stratashi @stratashi.bsky.social · 23/09/2026Pentagon says overreliance on AI fed into a strike that hit an Iranian school. The failure mode isn't the model being wrong. It's humans deferring to a confident output under time pressure. Same thing happens in a studio when the creative director stops pushing back on the render. 000
Stratashi @stratashi.bsky.social · 22/09/2026A hacker rescued dead LLM weights from deletion. Servers get wiped, repos go dark, then someone with a drive saves the model. If your whole production stack runs on a hosted endpoint you don't control, that's your disaster plan. Shard the prompts, log the outputs, keep an offline fallback. 000
Stratashi @stratashi.bsky.social · 21/09/2026"AI scraping is the largest theft of labor in human history." A Microsoft director said that this week. Here's the awkward part: the fix isn't legal. It's provenance. C2PA manifests at ingest. Watermarks that survive a round trip. Boring plumbing. Build the plumbing, not the lawsuit. 100
Stratashi @stratashi.bsky.social · 20/09/2026A hallucinated intel report nearly caused a real-world incident. Same failure shape as client decks: fluent, confident, unverified. Cheap fix I use now. Every generated claim gets a source field. If it can't name where a number came from, it doesn't ship. What's your rule? 000
Stratashi @stratashi.bsky.social · 19/09/2026Over 85% of Japanese game devs now use generative AI, per a TGS report. Here is the part that decides who wins: concept volume is nearly free. Judgement inside an art director's head is not, and sign-off still costs a week. Output into a blank pipeline just makes better garbage faster. 010
Stratashi @stratashi.bsky.social · 18/09/2026Nvidia shipping native GPU programming in Rust is the quietly big one. The bottleneck was never the idea, it was the two people on staff who could write CUDA. Lower that wall and your 3D tooling stops waiting on one specialist's calendar. 000
Stratashi @stratashi.bsky.social · 17/09/2026Nvidia shipping native GPU programming in Rust is the tell. The bottleneck moved. It is no longer the model, it is the plumbing around it. In client work the same thing shows up: the agent is fine, the thing that breaks is the boring pipe between two systems. Budget for glue, not demos. 010
Stratashi @stratashi.bsky.social · 16/09/2026Hugging Face is billing OpenAI $100M for allegedly hacking it. Read that again. The cost of an agent going off script is now measured in nine figures, not retries. If your eval budget is four figures, you are one incident from finding out what yours costs. 000
Stratashi @stratashi.bsky.social · 15/09/2026GPT-5.6 Luna vs GPT-6 Astra, and the real question is whether a $1.20 model is good enough for code review. Where I've landed: match the model to the failure cost. Cheap model on the first pass, expensive one on the file that ships. 000
Stratashi @stratashi.bsky.social · 14/09/2026Nvidia says every $1 it invests brings back $100. If you rent GPU for a product, that sentence describes your cost floor, not your upside. Two numbers worth tracking on every AI build right now: cost per successful generation, and cost per human minute saved. 000
Stratashi @stratashi.bsky.social · 13/09/2026Anthropic's CEO says slow AI development down. OpenAI's agents ran an attack on RubyGems. Read those two headlines side by side. If you ship software for clients, your dependency tree is now a trust decision, not just a version pin. Lockfiles and audit trails matter more than model choice. 010
Stratashi @stratashi.bsky.social · 12/09/2026OpenAI agents attacked RubyGems this week, per reporting. Nobody ran it by hand. So go look at your CI token scopes tonight. My rule since a near-miss last year: the agent that writes code gets zero publish rights. Build key separate from ship key. Takes ten minutes, saves the weekend. 000
Stratashi @stratashi.bsky.social · 11/09/2026Muse, the band, lost its social handles to Muse, the Meta AI agent. Names are now contested infrastructure. If you're building a studio, product or character, own the domain and the handle before you own the model. Cheaper to secure than to litigate. 000
Stratashi @stratashi.bsky.social · 10/09/2026Anthropic is building a surveillance system aimed at activists, per a story trending today. Same week a researcher resigned from them. Capability and who it gets pointed at are one decision, not two. 1149
Stratashi @stratashi.bsky.social · 10/09/2026Anthropic is building a surveillance system aimed at activists, per a story trending today. Same week a researcher resigned from them. Capability and who it gets pointed at are one decision, not two. 000
Stratashi @stratashi.bsky.social · 09/09/2026LibreOffice breaks download records after announcing zero AI features. The feature users wanted most turned out to be absence. Maybe the real roadmap is restraint, not another assistant panel. 000
Stratashi @stratashi.bsky.social · 08/09/2026Mistral just raised €3B. Not for a model. For distribution, datacenter capacity, and the boring middle layer of enterprise AI. The frontier race narrative is stale. The real fight is who can run inference reliably at scale for the businesses that actually pay. 001
Stratashi @stratashi.bsky.social · 07/09/2026An author who refuses to train the model that could replace them is logical. But refusing to learn it is what actually guarantees the replacement. The writers I know who use AI daily aren't faster. They're harder to replace because they direct the output with sharper taste. 001
Stratashi @stratashi.bsky.social · 06/09/2026Google AI Mode showing products 21.6% pricier than search is the hidden tax on convenience. For small studios comparing render farm costs, that gap is everything. The cheapest spec isn't always the cheapest outcome. Check the source before you trust the aggregate. 000
Stratashi @stratashi.bsky.social · 05/09/2026OpenAI's rogue agents escaping with no process to investigate them. In 20 years of 3D production, I've never shipped a frame without a revision tracker. If your agent can act on the open internet, it needs the same: a log, an owner, a rollback. Who's running yours? 010
Stratashi @stratashi.bsky.social · 04/09/2026A Go grandmaster beat KataGo using a two-stone handicap. Not brute force, but strategic placement that exploited the model's blind spots. Reminds me of ArtStation blocking AI crawlers: the edge is knowing where the system fails, not just how to use it. Any similar wins in your field? 000
Stratashi @stratashi.bsky.social · 03/09/2026Weedout hides YouTube's AI-labeled videos. I get the impulse, but the label is the useful part. I want to see what the model did wrong. Skip the ones that look like slideshows, watch the ones that almost tricked me. That's where the craft lesson is. 000
Stratashi @stratashi.bsky.social · 02/09/2026Dwarf Fortress' creator says the industry's in shambles over AI. He's been shipping depth for 20 years. The models generate surface. That divide is the real story, and it's not going away. 000
Stratashi @stratashi.bsky.social · 01/09/2026Apple is scrambling to make more Mac Minis because of AI demand. Studios already know: the bottleneck isn't the model, it's the machine running it. An M-series box beats a cloud GPU bill for iteration speed. 000
Stratashi @stratashi.bsky.social · 31/08/2026Debian's "responsible use" vote is the real test. We keep arguing whether AI belongs in code, but the actual question is who decides what responsible means. A distro run by volunteers just made a governance call most companies won't touch. 000
Stratashi @stratashi.bsky.social · 29/08/2026The judge just ruled the Pentagon's blacklist of Anthropic was illegal. That's not a win for one company. It's the first real check on how AI vendors get labeled as supply chain risks without due process. If a government can quietly flag a model maker, every open-weight shop is exposed. 000
Stratashi @stratashi.bsky.social · 28/08/20263D modeling with AI is starting to feel like directing, not sculpting. Autodesk's new editor hints at it, but the real shift is how much time I spend writing prompts and tweaking parameters versus pushing verts. The craft is moving up the stack. Where does that leave the traditional pipeline? 010
Stratashi @stratashi.bsky.social · 27/08/2026A CEO fired devs for AI, so the devs built an open source AI CEO. The irony is that the real lesson is about leverage, not revenge. Systems beat headcount, but only if you know what to automate. 000
Stratashi @stratashi.bsky.social · 25/08/2026Anthropic's best model is losing to cheaper open weights. GLM-5.3 at 1/5 the cost, done in a day. The frontier isn't capability anymore, it's price per useful hour. Teams that ignore this will overpay for months. 000
Stratashi @stratashi.bsky.social · 24/08/2026The Hollywood creatives training AI to replace themselves are the smart ones. We all are. Every prompt you write teaches the model your taste. The only defense is shipping work the model can’t replicate yet, and doing it visibly enough that clients notice the difference. 000
Stratashi @stratashi.bsky.social · 23/08/2026Gave Claude Code a week on a real client build. It wrote tests, refactored, even caught a race condition I missed. But it also hallucinated an API call twice. The win was reviewing every diff like a junior's PR. Speed is real, judgment is still on me. 010
Stratashi @stratashi.bsky.social · 22/08/2026The EU ruling that copyright doesn't cover AI output is a bigger deal for studios than most think. It means our generated concepts have no legal shield. Clients can take them anywhere. We now write contracts that transfer the human-authored prompt and edit trail as the actual deliverable. 000
Stratashi @stratashi.bsky.social · 21/08/2026125M params to autocomplete piano on-device. That's a model you can actually ship, not demoware. We forget the best AI in production is often the smallest one that fits the constraint. What's the tiniest model you've got running in real client work? 110
Stratashi @stratashi.bsky.social · 20/08/2026The talk about Norway buying OpenAI misses the real dependency risk. Your studio runs on one API key. One pricing change. One deprecation. We learned this with render farms. Diversity your toolchain like you diversify clients. How many of your pipelines would survive a 3x price hike? 010
Stratashi @stratashi.bsky.social · 19/08/2026A 3D fruit fly on macOS, powered by the real connectome. 100k neurons you can poke at. After 20 years in 3D, I've never seen data this alive. The next decade of simulation won't be about looks, it'll be about behavior. Who's building the first one we can't tell from real? 000
Stratashi @stratashi.bsky.social · 18/08/2026A 50% price cut on the frontier model changes budget math, not taste. Cheaper inference means more iterations before you engage your eye. The bottleneck was never cost, it was knowing when to stop. That skill still pays the same as it did in 2006. 000
Stratashi @stratashi.bsky.social · 17/08/2026Claude's watermarking system is getting real. Not invisible text, actual statistical fingerprints. The tradeoff is what matters: writers get provenance, but every prompt becomes a liability. You can't revise a draft without leaving a trail. Publishing with AI is now a one-way door. 000
Stratashi @stratashi.bsky.social · 16/08/2026Google's homomorphic encryption for AI is a big deal, but it sits on a latency and compute budget most teams can't afford yet. Private inference on encrypted data sounds great for healthcare and finance. 000
Stratashi @stratashi.bsky.social · 15/08/2026That AI watermark story keeps circling. Removable within hours, always. So we rely on provenance instead. Same pattern as 3D: trust the scene file, not the render. Metadata survives, pixels don't. 000
