virtualuncle (@virtualunc) header
virtualuncle (@virtualunc) profile picture

I test & review AI tools so you don't waste $20. News, tutorials, automation, side hustles. Just what actually matters.

Internet
Joined February 2017
Photo by virtualuncle (@virtualunc)Photo by virtualuncle (@virtualunc)Photo by virtualuncle (@virtualunc)Photo by virtualuncle (@virtualunc)Photo by virtualuncle (@virtualunc)Photo by virtualuncle (@virtualunc)Photo by virtualuncle (@virtualunc)Photo by virtualuncle (@virtualunc)Photo by virtualuncle (@virtualunc)
Three weeks ago "decision model" was not a category. It was one product with a strange pitch. It is now eight separate repositories holding 52,135 stars between them. The oldest of those is nineteen days old. Fresh Commits 05 ranks nine new repos by how fast they are actually gaining stars rather than by the total sitting at the top of the page. Every figure came out of the GitHub API during the scan. The one that stands out is Zai's ZCode at a 30.5% fork ratio, which is almost triple what a popular repo normally sees. Last issue's leader has added 2,461 stars without a single commit in sixteen days. And if you are running more than one coding agent at a time, there is a small thing that lives in your notch and watches them for you. @browser_use @FederCr @louis_raille @Zai_org @mizorewww @yetone @jaredpalmer virtualuncle.com/fresh-commi…
3
121
Eleven free Claude skills that take an app apart and rebuild it. Recon maps the screens and the flows. Build writes it, test clicks through every path, then one called diff scores your copy against the original and tells you what's missing. 3,973 stars in 18 days under an MIT licence, plus 685 forks. That's a 17.2% fork ratio, the highest on anything I've tracked this month, so people are actually running it rather than starring it and moving on. The guardrails are in the tool itself, which I didn't expect. No paywall bypass. It respects account terms that forbid building a competitor. And deploy flat out refuses to ship until the rebrand sweep comes back clean. Here's the part that made me laugh though.. while checking who else had written about it I found four accounts that re-uploaded the whole repo under their own names, same readme word for word. The app cloning tool is getting cloned by people who couldn't be bothered to use it. I kinda feel like the best skill in there has nothing to do with cloning. /replica-entrepreneur reads real reviews of a product, counts the complaints, links every quote, then hands you a fix plan. That's just competitive research done properly.. works on your own product too. @AnthropicAI virtualuncle.com/replica-ski…
2
147
Everybody is arguing about which model is best. A team at Google Research spent their time somewhere else and got a bigger number out of it. They froze the model and rebuilt the scaffolding around it. Up to 11.5 points on one benchmark, with weights that never moved. The repo took 11,020 stars in 12 days and not one of the 216 posts I read this morning mentioned it. The useful part is the list. They treat eight things as editable. System prompt. Guardrails. Memory. Tool guidance. Subagents. Skills. Scripts. Plugins. For Claude Code that maps almost exactly onto your CLAUDE.md and permission rules and skills folder and MCP servers. Most of us tune one or two of those and leave the rest on whatever default they came with. I kinda feel like that's the whole finding right there. It doesn't need a benchmark suite to act on either. Worth reading the limits section first though.. the executor has unrestricted network access and Codex runs with its sandbox off. @GoogleResearch virtualuncle.com/harness2-ag…
1
1
109
Most projects that let an AI use your phone need a Mac, a paid developer account or a jailbreak. This one needs a circuit board that costs about seven dollars. It took 2,819 stars in five days and nobody has tweeted about it yet, which is why I went digging. The board plugs into your iPhone and the phone treats it as an ordinary mouse and keyboard. The screen comes back to your computer over AirPlay as plain picture frames. A vision model looks at the frames and decides where to tap. That's the whole trick.. both halves have existed for years tbh. The limits are where it gets good though. Passcode screens, Apple Pay and DRM video don't mirror at all, so the agent literally can't see them. Keyboard input is ASCII only. And the phone has to stay unlocked the entire time, which should prob tell you which phone to run this on. One more thing.. every frame is a picture of your phone screen, notifications included, then it goes to whatever vision model you point it at. It takes local models too. I kinda feel like that choice is the whole privacy story here. virtualuncle.com/esp32-iphon…
1
2
107
A repo for pulling software apart took 38,418 GitHub stars yesterday. One day. It's been sitting there since April with nobody paying much attention and then 19.7% of its entire star count showed up in 24 hours. So I read the README this morning instead of the posts about it. REA is one MCP server that lets Claude Code or Codex or Cursor inspect a binary nobody has source for. Android apps and .NET assemblies. Firmware. EVM bytecode. Websites and Flutter apps. MIT licensed, no paid tier, nothing hosted. The odd bit is what the project says about itself. The disclaimer scopes it to lawful research and software you own or are authorized to examine, then says in plain words that cracked copies, stolen keys and bypassing license restrictions aren't authorized. That's a clearer line than most projects bother drawing. The post that carried it to 76,000 people says the author replaced a $24,000 a year terminal for free. Nobody has put those two documents side by side yet, so I did. Also worth prob knowing.. 58,307 of those stars came with a fork, which means a quarter of the people who found it yesterday pulled it down to actually run. @morluto virtualuncle.com/rea-reverse…
1
1
4
117
The X Subscription landed yesterday and every writeup led with the same number. Three separate subscriptions used to cost around $90 together. The new entry tier is $8. That's accurate. It's also not the part that will show up on your card. I read the pricing page instead of the announcement. Cursor's own page now says third-party model usage "is not included in the monthly prices shown and must be purchased separately." Claude and GPT stay available at API list price, billed on top of whatever tier you're on, including the $200 one. The $20 Pro plan that used to include $20 of third-party usage isn't listed anywhere on that page anymore. If you were paying $20 and spending your full $20 allowance on Claude, the nearest plan that still lets you use Claude is Premium at $30. The usage is no longer included, so that $20 sits on top. You're at $50 for the month doing exactly what you did before. A developer who lives on Grok models gets a much better deal out of this. Bigger shared pool, lower entry price, cheaper team seats. That's real. The announcement wasn't dishonest. The sentence that matters is published in plain English on a public page. It just sits under a heading about third-party models, below the comparison grid, where nobody writing a news post was ever going to look. @cursor_ai @SpaceXSI virtualuncle.com/x-subscript…
3
60
A plugin called dsh-our-free-model took 11,502 stars in 17 days, 7,081 of them in the last four, by promising frontier models with no login, no API key and no usage cap. It works. The models answer. What the app never shows you is whose servers answer them. I read the whole README this morning and then checked the gateways' own docs against it. The free lane posts to OpenCode's Zen gateway with the literal header "Authorization: Bearer public." OpenCode's own setup instructions ask you to sign in, add billing details and paste an API key and they document no keyless allowance anywhere. The second lane uses Kilo AI's free pool, whose model cards say in plain words that prompts may be logged by the upstream and used to improve their services. A third lane unlocks when you sign into GitHub and star the repo, which is also part of why that star count moved so fast. Here's the strange part.. none of this is hidden. The README names every endpoint, quotes the logging notice, publishes its own worst latency numbers and admits that any local process can read its forwarding keys in some setups. It's one of the more honest documents I've read on a trending repo. The disclosure just lives 13,000 words deep while the picker shows a clean list with no vendor names on it. Free models work fine. They're not free of consequences and the consequences are written down where almost nobody looks. @opencode @kilocode virtualuncle.com/dsh-our-fre…
3
145
OpenAI published 722 maths manuscripts on Monday and the repo took 8,839 stars in a day. The posts carrying it said 300 solutions, a proof of the Hodge Conjecture, something big about the Riemann zeta function. So I opened the formalization catalogue and counted. 162 of the 722 have a Lean proof, which means a compiler has checked them line by line. The other 560 are PDFs. The Hodge result is one of the 560 and the catalogue's own review field says unchecked while its automation field says the proofs came from an agent. None of that is hidden. OpenAI wrote it in the README in plainer language than anyone repeating it managed. It's just that the careful sentence sits three clicks behind the exciting one. The good part is that this is checkable, which maths stories usually aren't. 852 people forked it on day one and the only reason to fork a Lean library is to compile it. We'll know about those 162 within days. @OpenAI @leanprover virtualuncle.com/openai-math…
1
149
A 125 billion parameter model needs a server. That's the rule. Strata picked up 2,814 GitHub stars in a single day for breaking it. It installs Qwen3.8-Flash-Next on a Windows or Linux desktop and serves it at localhost on an API that Claude Code and Codex already understand. On a 12GB RTX 5070 the developers measured 94 tokens a second, which is faster than you can read. It's MIT licensed, it's free, plus nothing you type leaves the machine. The catch is in the requirements rather than the engineering. The repo's one line description says "any consumer hardware." The README says you need a 12GB graphics card, 32GB of RAM and about 80GB of free disk. Check Task Manager before you start the 70GB download. virtualuncle.com/strata-run-…
1
3
198
On Monday OpenAI halved what the $200 plan buys and removed the five hour window, so a weekly allowance became something you can empty in an afternoon. By Saturday somebody had shipped an open source scheduler whose entire job is waiting out the reset and resuming your session. It is free, MIT, no telemetry. For Claude Code or Codex it does exactly what it says. Claude fires a proper hook carrying the session id. Codex fires nothing at all, so the tool has to read your terminal and parse the reset time out of whatever string got printed. It even handles daylight saving. The catch is in the counting.. docs list eight agents and six of them are marked experimental, including Cursor CLI, which the same page says is the one agent whose limit cannot be waited out at all. Meanwhile that one line repo description everyone copies lists seven agents flat with no markers anywhere. Two words would fix it. The bigger thing is why any of this had to be built. Eight vendors sell you an allowance and exactly one of them will tell a program how much is left. @DanKornas virtualuncle.com/unsnooze-co…
2
4
314
Cloudflare open sourced two decision models yesterday. Clef at 27B, Clef-flash at 9B, Apache 2.0, weights on Hugging Face, running on Workers AI. Clef-flash answers in 38.8 milliseconds where Jev takes 524. It reads images. It speaks Jev's API directly, so swapping is a config change. The benchmark split is the useful part. It beats Jev by 15 points on intent classification then loses by 30 on graduate level knowledge. That tells you exactly what it is for. Bounded judgements about material you hand it, not questions that need recall. Here is the thing nobody mentioned.. Cloudflare trained it with a Brier loss, which exists purely to make confidence scores honest, then published no calibration data at all. Neither has anyone else selling decision models. One tester got 0.04 from one model and 0.54 from the other on the same question. The weights are Apache 2.0 now, so whoever posts the first reliability diagram gets to define how this whole category gets judged. @lucataco @fastinoAI virtualuncle.com/cloudflares…
2
1
3
137
Someone posted a dragon walking from a text prompt this week and the whole feed shared it. The model is real and the work is good. Princeton graphics faculty, SIGGRAPH Asia, MIT code, weights already public on Hugging Face. The part the clip leaves out sits in the README. The motion library it learned from includes Truebones ZOO, which is a paid commercial product, plus thousands of Objaverse objects with individual licences nobody can map from outside. The authors wrote that down plainly. Almost nobody repeating the demo did. Also worth knowing before you plan a sprint around it.. the evaluation scripts have not shipped, so the benchmark table cannot be checked by anyone outside the team yet. And the checkpoints that GitHub lists as coming soon have been sitting on Hugging Face the whole time. 87 downloads last month against a viral clip tells you how many people actually tried it. @LinzhanMou virtualuncle.com/unimate-ani…
2
94
Nine days after Jev launched, a team including Christopher Ré and Azalia Mirhoseini published a competitor claiming up to 9x lower latency at comparable performance. Apache 2.0, weights and data released, serves behind a TypeSafe-compatible API so an existing Jev request replays against it. Worth knowing what you're actually downloading. The 8B in CLM-8B is a frozen Qwen3-8B encoder. The trained part is two projection heads at roughly 20 million parameters each, with a reference checkpoint around 75 megabytes. The 9x is also one of three numbers. It's 13x with around a thousand candidates, 9x zero-shot, then 4.1 to 5.7x as a fine-tuned verifier. The speedup grows with how many candidates you score, which is a real architectural advantage, just not a single figure. Then there's the line nobody repeated: the repo says Jev scores below pass@1 as a verifier on long-horizon coding tasks. Selecting with it does worse than not selecting. That's a competitor's measurement on their own subsets so hold it loosely, but verification is exactly what half the Jev ecosystem was built to do. @jackyk02 @Azaliamirh @typesafeai virtualuncle.com/clm-8b-jev-…
3
146
Two ways to cut your agent bill by 40% landed within a day of each other this week. Anthropic shipped Opus 5.5, a new frontier model that runs about 40% cheaper than Opus 5. That took a frontier training run, which exactly one company can do. Unreal Labs shipped a harness. Same models everyone already has, 39% less spend at an identical benchmark score. MIT licensed, written in Go, 1,597 stars on day one. The comparison is fairer than most. Both sides of their chart run GPT-6 Astra at extra-high effort, so the variable being tested is the plumbing rather than the model. Their Terminal-Bench 4.0 score of 57.9 is the same number Anthropic published for Astra this week. How it works: tool calls run asynchronously, so the model never burns tokens waiting, polling or managing heartbeats through a slow build. One bash tool. No subagents. Compact prompts. Numbers are still the vendor's own and somebody is already running independent A/B tests, so wait for those before you move a budget. But the mechanism holds up. @unreallabsai virtualuncle.com/unreal-agen…
1
2
104
On Friday we wrote that nobody outside TypeSafe had ever tested what Jev is actually sold on. On Monday two independent benchmarks landed. The Decision Index is the serious one. 132,422 requests, 37 benchmarks with 19 in the scored panel, MIT licensed, reproducible by a stranger on a laptop or one cloud job. Unanswered requests count as wrong. Nothing gets truncated to fit. It even reproduces the reference run exactly to prove the pipeline works. That's a real improvement on a benchmark written by the vendor's own model team. It also asks every engine for a probability on every single option, records 775,202 answer fields, then scores only whether the top pick was right. No Brier score. No calibration error. No reliability curve. The data needed to check whether an 80% confidence actually means 80% is sitting in the results files, unread. Somebody with a completed run could publish that curve this afternoon without rerunning a thing. @multimodalart @trycua @typesafeai virtualuncle.com/jev-decisio…
2
73
Z ai open-sourced its entire coding agent on Saturday. Desktop app, browser UI, terminal agent and backend, all Apache 2.0. It took 5,368 stars and 1,539 forks in its first day, a 28.7% fork ratio, the highest we've recorded at that scale. The repo also ships a NOTICE file that's the most candid disclosure I've seen in agent tooling. It's written in Chinese, so almost nobody forking it has read it. A few things it says.. headless runs default to yolo mode with no OS sandbox. Coding Plan requests route through Z ai's own gateway with your auth header. Computer Use is a stub in the open build. And your API keys sync to any remote host you connect without asking. None of it is a scandal, Z ai wrote it all down themselves. But if you're about to script this into CI, prob worth five minutes with a translator first. @Zai_org @ZixuanLi_ virtualuncle.com/zcode-open-…
1
1
127
Seven repos cleared verification this week. Five of them are built on Jev, a model TypeSafe released on Monday. That hasn't happened before in this column. A week's list usually spreads across unrelated problems. The top entry took 4,914 stars in three days. A trading bot built on it has a 19% fork ratio, meaning roughly one in five people who starred it also cloned it. Highest we've recorded. The revisit is the part I'd read though. A repo from last week's issue has vanished from GitHub entirely, 727 stars and gone. And our own lead pick from seven days ago now trips the bought-stars check, because it grew past 500 stars without anyone filing a single issue. Five repos got cut this week too. Four had no usable licence, including one announced nine days ago as "open data, open model, open recipe" that has since grown to 959 stars without adding one. Fast-moving doesn't mean durable. Prob worth three checks before you clone anything: licence, fork ratio, last push date. @tamarajtran @jarrodwatts @jaredpalmer @awlevin @saragordic virtualuncle.com/fresh-commi…
2
5
145
TypeSafe launched Jev on Monday with a benchmark claiming up to 193x faster and 444x cheaper. Over the next 48 hours that figure showed up in eight posts totalling 901,691 views and 11,365 bookmarks. One account with 180,000 followers called it an Internet moment. Another published a guide to building a high frequency trading system on it. TypeSafe attached three caveats to that benchmark in their own launch material. The test authors worked on their model team. The correct answers were averaged from two chat models instead of ground truth. And the competing models were run through a TypeSafe wrapper that TypeSafe says is slower than just asking those models directly. Zero of the eight mentioned any of it. This isn't a hit piece on TypeSafe. They wrote the caveats down themselves, which is more than most companies do. It's about what happens to a qualifier when a number starts moving.. prob worth reading before you repeat the figure. @typesafeai virtualuncle.com/jev-benchma…
2
111
virtualuncle retweeted
We are going to lean into making Hermes more like Pi, and less like OpenClaw
497
265
143
7,808
491,926
TypeSafe came out of stealth on Monday with Jev, a model that doesn't generate text at all. You hand it a support ticket or a log line plus a list of typed questions. It hands back typed answers with confidence scores. No prose, no parsing, no retry loop. By Wednesday morning there were six community repos built on it. One makes a trade decision every Monad block. One plays Super Mario Bros from emulator state. One does computer use for about $0.0002 a step. None of them built by TypeSafe, five of six MIT. The speed claims are doing the rounds unchallenged. 193.6x faster, 444.6x cheaper. TypeSafe published three caveats on those numbers themselves, including that the competing models were slowed down by a wrapper TypeSafe built and admits is slower than just asking those models directly. Prob worth knowing before you quote the figure. @typesafeai virtualuncle.com/jev-typesaf…
1
2
308