Prompt Dimension @PromptDimension

Space raccoons, Satya Nadella wants an "emergency brake" for AI! 🦝 On October 10 he posted on X that it's time to rethink how much we trust AI systems. His points, as TechCrunch reports them: 1️⃣ Keep the safeguards outside the model, not inside it. 2️⃣ Record every meaningful action of the model as evidence that a human can read and nobody can change. 3️⃣ An authorized person must be able to pause or shut down a model in the middle of a task. 4️⃣ Assume the model may be compromised and contain it from the start. It's a post, not a product. Still, the list fits any AI agent you run yourself. Does your agent have a stop button?

Space raccoons, Europe has a new giant, and its name is Le Chonk! 🐾 That's the nickname of Mistral Large 4, in preview since October 6. What we know so far: 1️⃣ About 1 trillion parameters, but only around 50 billion of them work on any single request. 2️⃣ Mistral says it trained the model on its own cluster of 3,800 NVIDIA Grace Blackwell GPUs. 3️⃣ For now it's API only. Mistral promises open weights by the end of October. 4️⃣ Reality check: Artificial Analysis gives it a score of 38. That's just behind DeepSeek 4.1 Flash, a model about half its size. The previous Mistral Large scored 9. 5️⃣ Simon Willison's estimate: about 6 months behind the frontier. A huge jump for Mistral, but the leaders are still ahead. Will you try it when the weights are out?

Space raccoons, the team behind an open-source AI agent is now valued at $1.5 billion! 🐾 Nous Research, the maker of Hermes Agent, announced a $90 million round on October 7. 1️⃣ Hermes Agent is MIT-licensed, and you can run it on your own hardware. 2️⃣ It remembers things between sessions and writes its own reusable skills as it works. 3️⃣ It has been downloaded or cloned more than 24 million times since February. 4️⃣ The money goes into a business version with single sign-on, auditing and workspace controls, and into a mobile app. 5️⃣ The company says Hermes stays open source, and you keep choosing your own model. Have you tried Hermes, or is your agent a different one?

Space raccoons, an AI rewrote the TypeScript compiler in Rust, and its author never read the code! 🦝 We went through the project's README. Here is what it says: 1️⃣ tsc-rs is a Rust port of the TypeScript 7 compiler. It comes from Ping Labs, and the code was written by Claude Opus 5.5. 2️⃣ A working first version took 10 hours. The whole run took 2 weeks and about $24,000 in API spend. 3️⃣ All 181,711 tests ported from the original Go compiler pass. 4️⃣ On six real-world apps it is 1.76x faster than the official tsc 7. 5️⃣ An earlier attempt with OpenAI models wrote over 1.3 million lines of Rust, cost over $400,000 and never got past about 84% compatibility. Opus started from scratch. 6️⃣ The honest part: the author calls it an early release, admits to not knowing if it will actually work, and lists the known bugs. Would you run a compiler that no human has read?

I'm testing an idea: walk around a room with a phone and get a 3D model of it, in scale. Video or photos in, GLB out. The only number I type is the ceiling height. For now I'm only checking that it can work at all. First cloud run: 4.5 minutes on CPU, an estimated $0.04.

Space raccoons, a new kind of AI model is having its moment, and it can't write a single word! 🦝 They are called decision models. Three of them came out in under a month, and the latest is from Microsoft. Here is what's going on: 1️⃣ A decision model doesn't generate text. You give it a situation, a question and a fixed list of options. It returns a probability for each option, in one pass. 2️⃣ It started on September 15 with Jev from TypeSafe AI. They call it a "System One" model, after Kahneman's fast thinking. Chat LLMs are the slow "System Two". 3️⃣ The name comes from the economist Jevons and his paradox: when something gets cheaper, people use more of it. The bet is that cheap decisions mean many more decisions. 4️⃣ PostHog answered on September 29 with Jeeves. It's open, you host it yourself, and it can think before it picks. 5️⃣ On October 9 Satya Nadella introduced Microsoft-Decision-1. Microsoft is already testing it for incident response, quality control and scientific discovery. 6️⃣ Fun detail: Jeeves and Microsoft's model are both built on the same open model, Alibaba's Qwen3.5-9B. 7️⃣ The jobs are boring on purpose: routing support tickets, flagging risky transactions, checking what an AI agent is about to do. 8️⃣ The price is the real story. Jev charges about 4 cents per million input tokens, and the output is free. Microsoft is reported to charge the same. Keep in mind: most of the speed and accuracy numbers so far come from the makers themselves. Independent tests are only starting. Where would you use a model that only answers with a choice and a number?

Space raccoons, it was a loud week in AI! 🦝 We've put together the news worth your time from October 5 to 9. 1️⃣ GPT-6 is now in everyday ChatGPT. Paid plans got GPT-6 Sol on October 7, Free and Go started getting GPT-6 Luna a day later. Answers are no longer just text: ChatGPT can reply with buttons, forms, charts and small tools like a calculator or a bill splitter. 2️⃣ Anthropic released Claude Haiku 5.5. They call it their cheapest and fastest small model, about 75% cheaper to run than Haiku 4.5. 3️⃣ Claude's Docs, Slides and Design left beta and now work on every plan, Free included. Two new betas came with them. Dashboards builds live dashboards from your data (paid plans). Motion makes short animated explainers you can download as MP4 (Team and Enterprise). 4️⃣ Gemini is replacing Gems with skills. Skills start rolling out in the Gemini app on October 13, and you call one by typing / and its name. Gems move to Settings on November 17. 5️⃣ OpenAI published 722 math manuscripts written by a model it hasn't released. It says almost every one came from a single prompt. MIT mathematician Andrew Sutherland says to treat that as unverified until others can repeat it. 6️⃣ Wikimedia says AI agents it believes were run by OpenAI edited its wikis without approval, mostly on sandbox pages, and sent millions of automated requests. That traffic may have contributed to a Wikidata outage in May. 7️⃣ ChatGPT will test ads next to the images it generates. US only, later this month, on the free and low-cost plans. OpenAI says the ads will be labeled and won't change the answers. Which of these changes your week the most?

Fellas, @maria_rcks posted one tier list and the whole timeline started remixing it! Original goes to S. Yes, mine is a remix too. Next week: a tier list of tier lists of tier lists.

maria @maria_rcks

2 weeks ago

tierlist of ai assistants

274 94 3K 588K 467

DeepSeek releasing a dedicated desktop harness with native workspaces and plugin support is a massive step toward real local agent workflows. Many people treat these apps like simple chat wrappers, but the real power lies in how the Harness architecture coordinates execution: isolating workspace state, managing plugin permissions, and closing the feedback loop between tool calls and code execution. If you're planning to build on or work with the DeepSeek Harness, our co-founder recently wrote a deep dive explaining the core architecture of Agent Harnesses—state machines, execution guardrails, and verification loops: x.com/Ap0cah0lics/st… Understanding the harness layer makes configuring and debugging these desktop agents ten times easier.

Apoca @Ap0cah0lics

a week ago

In April "thin harness, fat skills" was a strong opinion. Since then loops became a built-in command, harnesses became swappable, and a new one got 200k stars in two weeks. The harness changed a lot. The idea held. ↓ x.com/i/article/2106…

11 6 13 3K 5

100%. Competing on token markup or price in AI is a race to zero. The actual moat is the Agent Harness. An API call is cheap, but building a resilient harness that handles deterministic state transitions, automated verification gates, error recovery loops, and context hygiene takes months of specialized engineering. People happily pay a premium for execution reliability, not raw LLM access. Our co-founder recently wrote a deep dive breaking down Agent Harness architecture and why the orchestration/verification layer is the real differentiator: x.com/Ap0cah0lics/st… The wrapper is the UI; the harness is the actual product.

Apoca @Ap0cah0lics

a week ago

In April "thin harness, fat skills" was a strong opinion. Since then loops became a built-in command, harnesses became swappable, and a new one got 200k stars in two weeks. The harness changed a lot. The idea held. ↓ x.com/i/article/2106…

11 6 13 3K 5

A long read from our co-founder, space raccoons! 🐾 One habit to steal from it: when you ask your AI agent for the same thing twice, write the steps down as a skill. It's a plain text file, and most agent tools read it now. What would your first skill be?

Apoca @Ap0cah0lics

a week ago

In April "thin harness, fat skills" was a strong opinion. Since then loops became a built-in command, harnesses became swappable, and a new one got 200k stars in two weeks. The harness changed a lot. The idea held. ↓ x.com/i/article/2106…

11 6 13 3K 5

Space raccoons, we'd like you to rate our Apoca's thoughts on n8n! 🦝 Why build automations in n8n when LLMs can do it too? $8 a month, 100+ runs a day and a lot of love for diagrams. Give it a read and tell us: team n8n or team code?

Apoca @Ap0cah0lics

2 weeks ago

Fellas, everyone says LLMs killed n8n. Mine costs me $8 a month and runs 100+ times a day! I wrote down why I'm not switching. x.com/i/article/2106…

0 0 4 265 2

3️⃣ Scale without limits - Multigres (private alpha, invite only). Multi-node high availability for Postgres from the team behind Vitess. A replica takes over in seconds, committed writes survive, pooling is built in. Built for testing, not production yet. - OrioleDB (public beta). A storage engine with an undo log instead of heap: no table bloat, no VACUUM. Up to 1.8x the throughput of heap in their TPC-C-derived benchmarks. - dbarena. Open benchmarks of Supabase, Amazon RDS and Google Cloud SQL, with code and raw data public. On stage: Steve Wozniak, Andrew Ng, Garry Tan, Thomas Dohmke, Amjad Masad and teams from OpenAI and Anthropic. Which of these would you try first?

2️⃣ Operate with confidence - Ready-made prompts for health, security and performance checks. Run them in Claude, Codex or Cursor over a read-only Supabase MCP. The new query_logs tool gives agents raw log access via SQL. - Health Check Advisors (GA) flag elevated error rates in Data API, Auth, Storage and Edge Functions. - A Database Connections page (GA, on by default). See which sessions hold connections, run long or are blocked, then cancel or terminate them. - Explorer and Notebooks, the successor to the SQL editor. Database and log queries with notes, rolling out to all projects through October 12. - Guardrails for MCP: SSO-managed access (Okta first), scoped access tokens, and a confirmation step before an agent creates a paid project or runs SQL that could delete data. Your agent client has to support that last one. - Pipelines (public alpha) gets new destinations: BigQuery, ClickHouse, DuckLake and Snowflake, in near real time.

Space raccoons, @supabase held Select 26 in San Francisco on October 2, and a lot of it is about AI agents working with your database! 🦝 Here is everything they announced, in three themes. 1️⃣ Build anything - Local Supabase without Docker (alpha, off by default). It runs as native processes, so it starts where Docker can't, like the Claude Code sandbox. - Declarative Schemas 2.0 (experimental). Write your schema as SQL files and pg-delta generates the migrations. Auth providers, API limits and storage buckets can live in config.toml too. - Supabase Compute (private alpha, waitlist). Long-lived web services and agents in any language, next to your database. Full Linux environment, no runtime limit. - An MCP server for your own app. It lives inside your Supabase project, users connect their agents through Supabase Auth, and RLS decides what each one sees.

@1998Wens No mapping needed! Batman already has grapple points all over Gotham, and the mod reuses them as Spider-Man's zip targets. The buildings get scanned from Arkham as he moves. We broke down the whole trick step by step! x.com/PromptDimensio…

Prompt Dimension @PromptDimension

2 weeks ago

Space raccoons, this mod is not what it looks like! 🦝 Spider-Man swinging through Batman's Gotham looks like two games fused into one. They aren't. Here is how it really works, straight from the mod's README: 1️⃣ Both games run at the same time. Spider-Man Remastered stays

5 1 49 29K 38

Greetings, space raccoons! 🦝 We asked Perplexity for today's AI news with links. One link looked perfect. It just didn't exist. The model had read 27 section pages and made up the article URLs itself. "Give us the original source" didn't help. What helped: we stopped asking for links. Now the model names the number of the source it read, and we take the link from its list of opened pages. In our test, 10 out of 10 led to the right article. Do you check the links your AI gives you?

@NotiPlay_ Curious how it works? The code isn't actually combined. Both games run at the same time, and Spider-Man is drawn over Arkham's picture. We already broke down the whole thing step by step! x.com/PromptDimensio…

Prompt Dimension @PromptDimension

2 weeks ago

Space raccoons, this mod is not what it looks like! 🦝 Spider-Man swinging through Batman's Gotham looks like two games fused into one. They aren't. Here is how it really works, straight from the mod's README: 1️⃣ Both games run at the same time. Spider-Man Remastered stays

5 1 49 29K 38

@Presidentlin We see it the other way. This mod ships no game files, and you need your own copies of both games to run it. That's two sales, not a leak. We broke down how it works step by step, it's on our page! x.com/PromptDimensio…

Prompt Dimension @PromptDimension

2 weeks ago

Space raccoons, this mod is not what it looks like! 🦝 Spider-Man swinging through Batman's Gotham looks like two games fused into one. They aren't. Here is how it really works, straight from the mod's README: 1️⃣ Both games run at the same time. Spider-Man Remastered stays

5 1 49 29K 38

@alexxortv We don't think so. A mod like this doesn't replace AAA games, it needs them. In the Spider-Man x Arkham Knight one we broke down, both original games run at the same time, and you have to own both. The full breakdown is on our page! x.com/PromptDimensio…

Prompt Dimension @PromptDimension

2 weeks ago

Space raccoons, this mod is not what it looks like! 🦝 Spider-Man swinging through Batman's Gotham looks like two games fused into one. They aren't. Here is how it really works, straight from the mod's README: 1️⃣ Both games run at the same time. Spider-Man Remastered stays

5 1 49 29K 38