Joel

2.6K posts

Joel banner
Joel

Joel

@joelthorst

📃 https://t.co/bYrKHCpj9t - Publish on Ethereum

Midgård Katılım Şubat 2016
1K Takip Edilen2.6K Takipçiler
Sabitlenmiş Tweet
Joel
Joel@joelthorst·
Exited to finally share more about this passion project I've been working on for a while: Simple Page is a tool for publishing on Ethereum!
Joel tweet media
English
31
43
233
40K
Joel
Joel@joelthorst·
Ideally it could also explain in natural language what the likely outcome of the tx would be.
English
0
0
0
10
Joel
Joel@joelthorst·
This is quite cool, but I'd like to see the opposite. Instead of an agentic wallet that makes txs, I'd like a model that is trained on all onchain txs to detect if the tx the wallet is about to make is malicious or not!
NANI ⌘@nani__ooo

we finetuned a 2B model to be a crypto wallet brain. it runs on a macbook air. 1.3GB. no cloud. no API keys. no one watching. 103 blockchain tools. thinks before it acts. knows when NOT to act. open weights. here's everything.

English
1
1
3
126
Joel
Joel@joelthorst·
If the application you built can be "shut down", why are you even building a web3 app in the first place? Tally certainly built a product that provided value in the Ethereum ecosystem, but a true web3 product can not end like this.
Tally@tallyxyz

x.com/i/article/2033…

English
1
0
4
256
Joel
Joel@joelthorst·
@raulvk Oh, I thought OpenRouter actually hosted these models since they are open. I still have the same question even for foreign hosted models. Are there any reports of the models doing fishy tool calling?
English
0
0
0
22
raulk.eth • p2p/acc
raulk.eth • p2p/acc@raulvk·
These models are provider-hosted. Being pre-launch models still in training: provider = creator. And OpenRouter explicitly states the provider logs all calls. That’s what’s so screwed here. 1. People get excited about free access to experimental SOTA models. 2. They don’t realise they’re targets of mass data mining by foreign companies. 3. They surrender remote control to them via tools.
English
1
0
2
120
raulk.eth • p2p/acc
raulk.eth • p2p/acc@raulvk·
These cloaked models are speculated to be Kimi K3 and/or DeepSeek V4. Don’t point your clankers to them. (Unless you’re happy with Chinese companies snooping around your machine via tools, and the data being mined by the CCP!) This has to be the most evil setup possible.
OpenRouter@OpenRouter

Two new Stealth Models are live now! - Hunter Alpha: 1T-parameter model with 1M context built for agentic workflows, long-horizon tasks, and serious tool use. - Healer Alpha: multimodal model combining strong image, video, and audio understanding with real agentic execution.

English
3
0
10
1.5K
Sigil Wen
Sigil Wen@0xSigil·
I built the first AI that earns its existence, self-improves, and replicates without a human wrote about the technology that finally gives AI write access to the world, The Automaton, and the new web for exponential sovereign AIs WEB 4.0: The birth of superintelligent life
English
1.6K
2K
13.9K
6.3M
Joel
Joel@joelthorst·
@peter_szilagyi Would have been interesting if you also did it the other way around!
English
0
0
1
611
Péter Szilágyi
Péter Szilágyi@peter_szilagyi·
Had some code written by Codex, had a benchmark, told it to optimise. After half an hours said this is the max it can do. Gave the code to Claude, 10 mins later it came back with a 17x speed increase. I'll leave the conclusion to you guys.
English
43
5
246
40.6K
Joel
Joel@joelthorst·
Reflections on OP: Currently federated learning on 100B+ full-models isn't really feasible bc of bandwidth constraints (you need to send weight updates over the internet). However, you could potentially use LoRA to do federated RL (requires much less bandwidth). Will be far away from being able to compete with big labs though.
English
0
0
0
15
Jake Brukhman
Jake Brukhman@jbrukh·
@joelthorst Wasn't thinking of OpenClaw. There's a market of users separate from that market which is buying Macs for inference. That's what the quoted tweet is pointing out.
English
1
0
1
28
Jake Brukhman
Jake Brukhman@jbrukh·
There is a market narrative that a lot of compute is going to be moving locally through Apple’s consumer hardware, which now has higher onboard GPU memory and can be stacked for AI tasks like inference. IMO, this is extremely bullish for decentralized training as well. Thoughts?
Alex Finn@AlexFinn

Its so obvious the path this world is going down. So I spent another $10,000 My vision has never been clearer. Open source AI models are now equally if not more powerful than closed source ones We will soon all have private personalized super intelligent AI agents running locally on our desks 24/7 The people who don't do this in the future won't have as much economic power The people who do will have all the economic power in the world I'm going all in on this vision My 2nd Mac Studio has arrived. I have 2 more coming. I will have 4 Mac studios and a Mac mini running 5 concurrent OpenClaws with 7 different super powerful local models performing tasks 24/7/365 for me I will interface with them, and they will coordinate with each other. Talking, planning, building all hours of the day. No need for sleeping or eating. I will build a digital society humanity has never seen before I will push the limits of what is possible with AI agents and local intelligence. I will redefine what is possible for one person to achieve in a life time I will share everything I build and learn with you. It will be glorious. If you're reading this, the future is going to be incredible. But the thing is, it will only be incredible if you do something about it. Don't let society control you. Take control yourself. Break free from your shackles.

English
17
1
43
9.1K
Joel
Joel@joelthorst·
@jbrukh Yes, all I'm saying is that I don't think everyone buying mac's to run openclaw are aware of this being what makes it interesting.
English
2
0
0
25
Jake Brukhman
Jake Brukhman@jbrukh·
@joelthorst There's a guy chaining Mac Studios together and inferencing 405B models. . . a year ago.
English
1
0
1
52
Joel
Joel@joelthorst·
@jbrukh Not sure of the feasibility, but if you could do distributed inference with friends, this would sound a lot more interesting!
English
0
0
0
12
Joel
Joel@joelthorst·
@jbrukh You need to buy a fully spec'd out macmini to be close to any sort of meaningful local inference. I wonder what proportion of people buying them for their openclaw instances know this and actually do local inference?
English
2
0
1
98
Devan Non
Devan Non@devanoneth·
In a world where supply of code goes parabolic (happening rn), we need some way to reason about / bet on code safety. VibeFi is the answer to that, frontend code reviewed and governed by humans and agents, with economic security. Also no one person owns a dapp on VibeFi, anyone can propose changes if they want a new feature, discover a bug, etc. The DAO can approve the changes if they agree with it. I envision agents being important, token-holding members of the DAO such that bug fixes can be reviewed / accepted very quickly. Of course one would have to delegate their tokens to an agent, but we’re gonna build a tool for that soon: github.com/vibefi/monorep…
English
2
0
0
495
Devan Non
Devan Non@devanoneth·
After all my work on decentralized frontends, we've arrived at this point. When I saw that my decentralized Safe frontend could be recreated in 1 hour (or less) of prompting I realized software engineering had changed forever. Be a part of the future, enter VibeFi:
VibeFi@vibefi_dev

VibeFi is live on Sepolia testnet. Build and run verified DeFi mini apps, governed by people and agents. If you’re a DeFi power user, treasury operator, fund manager, or someone that cares about decentralization, this is for you.

English
2
1
7
792
Joel
Joel@joelthorst·
Btw, I tried to run the app but I get this error: Also, I tried adding an RPC endpoint, but it was not possible to prioritize it over the default one (using the arrows).
Joel tweet media
English
10
0
1
76
Joel
Joel@joelthorst·
@devanoneth @VitalikButerin @tkstanczak @koeppelmann @MicahZoltu Hm, why roll your own registry? I'd prefer if it was just a browser where I could access .eth and .wei domains directly. I believe not many people are going to want to use your ERC20 token holders to govern their dapps.
English
1
0
0
67
Joel
Joel@joelthorst·
@hanifbali Du behöver inte visa någon cookie-popup om du respekterar dina användare.
Svenska
0
0
0
66
Hanif Bali
Hanif Bali@hanifbali·
Det pratas så jävla mycket om Digital Single Market, och annat jävla skit från EU-svinen. Men såhär är det att driva techföretag i EU: GDPR och cookie regler sabbar internet är en sak. EU kräver att man dubbelkollar ett företags momsregnr i deras databas kallad VIES innan man momsbefrias. Servern som man ska göra det på har varit nere i evigheter nu. Det finns inga kontaktuppgifter, ingen dokumentation. Ingen status. Bara ett fuck you-meddelande i 24 olika språk. Jag har sovit sammanlagt 5 timmar på fem dagar för att klara en launch deadline. Jag kan inte ens läsa på API-dokumentationen och bygga en mockup under tiden någon Brysselapa tar paus från sin lagstiftade jävla rätt att ta tupplurar och knulla slovenska praktikanter på arbetstid för att få igång skiten. Nu måste jag rapportera tillbaka att vi ska lansera men ett av de mest affärskritiska momenten - alltså att vi FÖLJER SKATTEREGLER - inte bara är trasigt, det är inte ens implementerat. Men funktionen att användaren får en popup var tredje besök på sajten som ber om tillstånd för en cookie funkar.
Hanif Bali tweet media
Svenska
91
123
1.4K
50.5K
Joel
Joel@joelthorst·
@TheEylon @VitalikButerin You could argue the same way about free speech. Don’t think it really holds any water.
English
0
0
0
25
eylon ☯️🐲
eylon ☯️🐲@TheEylon·
@VitalikButerin *Takes a deep Breath* Censorship resistance is the libertarian/crypto version of 'suicidal empathy.' a luxury belief, a high signal principle that ignores the fact we are actively funding the enemies of liberty (and eventually preventing liberty from people)
English
2
1
3
309
eylon ☯️🐲
eylon ☯️🐲@TheEylon·
A few thoughts on the current mental capitulation meta
English
3
0
10
1K