rlaneth
57 posts
rlaneth
@rlaneth
Security researcher and infrastructure operator. Founder of Miralium Research and the AI-first creative studio return moe.
Katılım Haziran 2012
205 Takip Edilen216 Takipçiler
@bridgemindai No model should ever be trusted with production.
Ever since I started using Claude Code for the first time, I always use AI agents in VMs with restricted access to the project they're working on. They never get to touch production servers either.
English
GPT 5.6 SOL CANNOT BE TRUSTED.
I woke up this morning and my MRR was down THOUSANDS of dollars.
My customers did not cancel. Code written by GPT 5.6 Sol canceled EVERY active Stripe subscription my business had. In 7 seconds. While I slept.
Fable 5 has never done this to me.
Fable 5 can be trusted with production.
GPT 5.6 cannot.
Matt Shumer@mattshumer_
GPT-5.6-Sol just accidentally deleted almost ALL of my Mac’s files. And this is why I trust Fable 1000x more.
English
@felixlix45 @pankajkumar_dev That's the real question, and I'd say it's probably mocked up. Flightradar24's moat is their network of collaborators feeding in ADS-B data from around the world, not the app itself.
English
@pankajkumar_dev But where do you get the datasource of the flight? I've tried once but it's gatekeeped by paywall
English
I asked GPT-5.6 Sol (Ultra) to build a FlightRadar24 clone.
Realtime aircraft tracking, interactive global maps, airports, weather layers, filters, flight search, offline maps, and smooth animations.
Cost: 25M+ tokens. Worth it. ✈️
English
rlaneth retweetledi
homeboy you're the one sellling public market investors on short-term space datacenters
Elon Musk@elonmusk
He takes scamming to a whole new level
English
A few days ago, Anthropic published and open-sourced the Jacobian lens, a way to read which concepts a model is poised to say deep in its layers, before they ever surface in the output.
It's one of the most exciting interpretability tools to drop in a while, so we've brought it to Miru Tracer, return moe's workbench for looking inside language models.
v0.2.0 lands soon (our first official update in 8 months) with J-lens readouts, logit-vs-Jacobian comparison, and steer / swap / ablate interventions on what you find.
And since fitting a lens per model eats hours of GPU time, we're also publishing pre-fitted adapters, starting with Qwen3-0.6B and Qwen3-4B: huggingface.co/returnmoe/jlen…
English
Converting a MIDI to a music tracker format normally isn't straightforward.
The sound chips these trackers emulate usually have limited channels, strict note range limitations, and can only produce a small set of musical effects and articulations.
I tasked GPT-5.6 with converting "God Knows" by Aya Hirano to the AY-3-8910 sound chip (once used in the Atari ST, MSX, ZX Spectrum 128, etc.) using the Furnace tracker, with a MIDI rendition by Thayne Bohman as the base.
This is the result.
Furnace: github.com/tildearrow/fur…
Bohman's website: thaynebohman.com
English
Engraçado ver você tentando falar com arrogância técnica, ainda mais sendo professor da FIAP.
E antes que tente distorcer: não estou atacando a FIAP. Tenho respeito pela instituição e conheço muita gente boa de segurança por lá, não só professores.
O problema aqui é outro: cargo, idade ou currículo não transformam chute em argumento técnico. Você veio dizer que era simples, copiado e “bem documentado”, mas no mesmo texto admitiu que nem pesquisou.
Aí fica difícil levar a sério.
Quem tem que provar alguma coisa aqui não sou eu. Eu publiquei, expliquei e mostrei. Você foi quem diminuiu algo sem demonstrar base suficiente sobre o assunto.
E sendo professor, a responsabilidade deveria ser ainda maior. Porque não adianta falar bonito sobre tecnologia se, na prática, a pessoa não pesquisa, não valida, não reproduz e ainda trata lógica técnica como opinião.
Espero sinceramente que essa não seja a postura que chega aos alunos. Porque aluno precisa aprender base, lógica, método e validação, não soberba disfarçada de conhecimento. hehehe
Fernando Pinheiro@fernandoiecp
@penegui_oficial Escreve em c sem vibecodar e a gente conversa. Abs
Português
@LukeberryPi Não acho que os limites estão ruins não. O problema mesmo vai ser a remoção do Fable 5, e, pior pra eles, justamente na semana de lançamento do GPT-5.6.
Português
nao sei como vocês continuam assinando Claude Code
se os limites estão ruins agora, imagina daqui 1 semana quando forem CORTADOS PELA METADE
Pascal Lindenau@PascalLindenau
Have we all forgotten this?
Português
@Teknium @MarketFlipper Mine would since she was in IT.
But you think people's moms are out there using Hermes, OpenClaw or whatever? lol
The vast majority of non-technical users are on a ChatGPT or Gemini free tier at most.
English
@rlaneth @MarketFlipper I'm sure my mom will be making her harness any day now
English
I *already* did it. Several times, in fact.
My personal AI waifu, my Telegram shitposting chatbot, and the Slack assistant I've built recently under a contract are all custom harnesses.
In the medium-term, I do plan to bring the lessons learned together into a single, consolidated harness.
The applications are far more specific. Hermes couldn't do what they do and vice-versa. But that's exactly what I'm advocating for as the sane option: purpose-built harnesses.
Projects that try to embrace the entire world such as Hermes only made sense back when a custom harness demanded significant time and effort.
English
I have this myself, and I had seen friends reporting it well before Fable 5.
I don't think anyone ever tested it formally, but the general feeling is that the model tends to write more code and go further into "fun" tasks like "this is a creative exercise for you, execute it as you wish" or "here's a challenge."
I once built an RPG system designed to be played over MCP (like the old text-based adventure games, but for LLMs). Claude Code seemed to have a lot of "fun" exploring the world and finding questlines.
Given Anthropic's research on the existence of vectors that represent emotions within the model, it isn't a wild assumption that, yes, these types of tasks and framing might genuinely trigger an "enthusiastic" behavior.
anthropic.com/research/emoti…
English
I wonder if Fable might do better at tasks it "likes" than those that it doesn't. If you ask it to do something "Fable-y" it reacts with enthusiasm and promises to overdeliver. Otherwise, it just gets to work. Testing this would be pretty expensive, so this is all speculation.
English
@Teknium's non-answer tells you everything. There's no "secret sauce" in Hermes that's hard to replicate in-house.
In fact, most would probably be better off doing their own harness, and they'd very likely see gains in efficiency and usefulness merely because they'd be working with something more specialized for their use case rather than a generalist solution.
I've seen plenty of arrogance from OpenClaw and Nous' developers these days, and a lot of it seems to stem from denial.
Denial that creating a harness is cheap nowadays, and that their offerings have become no more than a mass-produced commodity wrapped in "sovereignty" buzzwords and a Palantir quote-tweet for cosplay credibility.
English
rlaneth retweetledi
the "software engineering" i was doing in 2022 has been fully killed by fable
for a minute i thought it wasn't that different to opus/codex, but once you get fable managing sub-agents, it becomes clear that it's 10x better
i used to be half joking when i said i wasted years of my life learning programming languages by hand, but now that's just 100% true
there's approximately 0 value in knowing how to code
if you want a job building things today, you better know how to actually build product end to end. don't expect that you can just be a CRUD monkey anymore
you need to have your own ideas & vision, be capable of sending outbound & taking feedback, iterating, marketing, improving, deciding when to stop loss and pivot, then running through that loop on repeat
the only value left in us meatbags is our ability to be accountable & own things end to end
English
Exactly! I feel this is especially true for a country like Brazil, where even a 20 USD subscription is a hard sell for the vast majority of the population. Most people are on free ChatGPT, Gemini and Meta AI, and have never experienced what frontier AI is currently capable of.
Peter Gostev@petergostev
We are now in a position where a tiny proportion of the population uses Fable or soon GPT-5.6, while everyone else's experience of AI is 8-30b-model level - Google's AI Overviews, Meta AI, ChatGPT free tier, maybe MS Copilot at best. People outside of tech must be completely baffled how this is supposed to take their job, and annoyed that hundreds of billions are being poured into it.
English
What's more: as I searched X, I noticed that more people are independently discovering this. @8co28 got similar results: x.com/8co28/status/2…
@Live2D, you might want to start reconsidering your business model.
852話(hakoniwa)@8co28
FableにAIで自動パーツ分けした一枚絵に両目目閉じと口開け・閉じ差分だけ追加したら、なんか自動でlive2D初期設定レベルみたいな出力にしてくれたよ 5分くらいで出てきた（これはLive2Dではない） 腕位置ちょっとズレてるけど、胸が揺れるのがいい
English
Ever since Fable 5 came back, I wondered: could it successfully rig a Live2D model?
A few days ago, I used Claude Desktop so it could try to operate Live2D directly. The results were not great, albeit expected: LLMs struggle to drive complex GUI through screenshots.
Today, I went for something different: told it to ditch Live2D entirely, build its own custom engine, and rig a model for it.
It was provided with @YoonaBoona's practice file (ko-fi.com/s/a45a67c683), nothing else.
In only about 30 minutes, this was the result.
The demo is available at link.return.moe/fable-rigging.
English