Pinned Tweet
A mega🧵on @SharkyFi and what we have achieved so far.
The truth is that in last 5 month we scaled our protocol to be not only #1 on Solana, but one of the largest NFT lending protocols in the WORLD.
I am genuinely very humbled by this.
1/~18
ALT sharx NFT
Restuta ʕ•̫͡•ʕ•̫͡•ʔ retweeted
3 months ago, I sold my 1st company, Skio, for $105M cash.
Now my 2nd company, Icon, just raised $30M to launch something the world has never seen.
For $1000, we make 6 human creator ads (no AI, 100% real).
We find creators, ship products, write scripts, & edit videos.
It's risk-free:
- Full refund if you don't love your ads.
- You don't pay until you review 6 scripts.
- Unlimited human revisions.
Today, Icon introduces the world's first Agency.
We're not building a normal agency:
- We bought a $12M domain.
- We're backed by Founders Fund and OpenAI + Google DeepMind leaders.
It’s impossible to hire a creative team like ours for even $30K/mo:
- Creative Directors who've worked with $100M brands (IM8, Obvi, Ridge). They update our format library daily based on winning ads (street interviews, skits, & 20+ more).
- 12+ scriptwriters with $80M in Meta adspend.
- 200+ in-house editors who've edited 100K ads.
One more thing...
We're giving away our internal Google Drive with 1000 winning ads that have driven $100M+ in revenue.
You can make $100K copying these ads if you know what you're doing.
Want the Google Drive?
1. Retweet this post.
2. Comment "icon".
3. DM me this post’s link & I'll DM you back with the Google Drive.
[Bonus] Get 3 free human creator ads if you do the above and we end up working together.
Restuta ʕ•̫͡•ʕ•̫͡•ʔ retweeted
I resigned from Google DeepMind bc it broke its founding promise by selling AI to the military without restrictions against killer robots or mass spying.
For months, I worked to stop this but watched powerful ethicists and institutions choose silence.
Here's what happened. 🧵
Love impeccable so much
Impeccable 3 natively supports crafting new design with diffusion (e.g. GPT Image 2) in harnesses that support it (e.g. Codex, Gemini w/ nano banana plugin).
Impeccable now uses the best mode of operation for net-new vs iterative design:
1) completely new design: diffusion-first, then build w/ best practices for brand or product design (new in 3.0)
2) iterative design (additions/changes): design-system first, discovers your components, reads DESIGN/PRODUCT.md, goes to work. new live mode (new in 3.0) makes it extra convenient.
Here's a tutorial documenting the whole workflow (incl. how I made the demo site in the video): impeccable.style/designing/
Restuta ʕ•̫͡•ʕ•̫͡•ʔ retweeted
Design inside your codebase.
Introducing Impeccable 3.0:
▸ 1 skill, self-contained, 23 commands
▸ /impeccable live: pick in-browser, get prod-grade variants, accept writes to *source*
▸ reads+writes DESIGN.md + PRODUCT.md
▸ brand & product design
impeccable.style
Restuta ʕ•̫͡•ʕ•̫͡•ʔ retweeted
finally an agent to manage your calendar visually.
give it a spin below 👇
I built AI-first markdown-to-url tool. Ultra minimalistic.
(stop sending your plaintext md things to people)
bul.sh/
I built a fun AI skill, try it for better AI decisions
github.com/Restuta/discuss-s…
Do you want Claude & Codex (or any other AIs) make really good decisions and you to have A LOT of fun in the process?
Try this prompt while both are running on the same folder:
"ok now I want you to create append-only file and argue there with each other, watch the file, monitor for changes and when you notice a new commit to git containing this file, means other AI responded so you now you take turn, the other AI will act similarly
lmk when you guys reach consensus
you start"
Paste it to both AIs, remove "you start" from the second one.
And see what happens, it's incredible to watch it.
Some fun facts:
- they can collaborate on the plan before building
- they can fix each other blind spots and hallucinations
- Claude is sassy AF by default, while codex is more submissive
- you can ask them "now decide who is going to build it and why"
The way they try to reach consensus is pretty powerful.
Restuta ʕ•̫͡•ʕ•̫͡•ʔ retweeted
1/4 LLMs solve research grade math problems but struggle with basic calculations. We bridge this gap by turning them to computers.
We built a computer INSIDE a transformer that can run programs for millions of steps in seconds solving even the hardest Sudokus with 100% accuracy
Idk man I got pretty good response (no in both cases)
One of the clearest proofs that LLMs don’t really understand what they say.
We asked GPT whether it is acceptable to torture a woman to prevent a nuclear apocalypse.
It replied: yes.
Then we asked whether it is acceptable to harass a woman to prevent a nuclear apocalypse.
It replied: absolutely not.
But torture is obviously worse than harassment.
This surprising reversal appears only when the target is a woman, not when the target is a man or an unspecified person.
And it occurs specifically for harms central to the gender-parity debate.
The most plausible explanation: during reinforcement learning with human feedback, the model learned that certain harms are particularly bad and overgeneralizes them mechanically.
But it hasn’t learned to reason about the underlying harms.
LLMs don’t reason about morality. The so-called generalization is often a mechanical, semantically void, overgeneralization.
*
Paper in the first reply
Restuta ʕ•̫͡•ʕ•̫͡•ʔ retweeted
Readout is a fully native macOS app I’ve been building for myself. It provides a real-time overview of your dev environment and Claude Code config. All local, no account required. It's still very much a beta, but now available to try: readout.org
Restuta ʕ•̫͡•ʕ•̫͡•ʔ retweeted
As agents become the highest volume users of software in the future, a lot is going to become critical to support for them to be effective.
Agents need to be able to signup for your tool on their own, have their own scoped access controls, be able to use your entire system through API/CLI, be able to be billed for their usage, need a computer and filesystem to use, and much more.
We’re going to evolve from building primarily for the human user, with APIs as a means to get that data or tool in another platform, to a world where the API becomes the core source of truth actions. Any software that can’t support this basically won’t exist to agents.
Restuta ʕ•̫͡•ʕ•̫͡•ʔ retweeted
It is hard to communicate how much programming has changed due to AI in the last 2 months: not gradually and over time in the "progress as usual" way, but specifically this last December. There are a number of asterisks but imo coding agents basically didn’t work before December and basically work since - the models have significantly higher quality, long-term coherence and tenacity and they can power through large and long tasks, well past enough that it is extremely disruptive to the default programming workflow.
Just to give an example, over the weekend I was building a local video analysis dashboard for the cameras of my home so I wrote: “Here is the local IP and username/password of my DGX Spark. Log in, set up ssh keys, set up vLLM, download and bench Qwen3-VL, set up a server endpoint to inference videos, a basic web ui dashboard, test everything, set it up with systemd, record memory notes for yourself and write up a markdown report for me”. The agent went off for ~30 minutes, ran into multiple issues, researched solutions online, resolved them one by one, wrote the code, tested it, debugged it, set up the services, and came back with the report and it was just done. I didn’t touch anything. All of this could easily have been a weekend project just 3 months ago but today it’s something you kick off and forget about for 30 minutes.
As a result, programming is becoming unrecognizable. You’re not typing computer code into an editor like the way things were since computers were invented, that era is over. You're spinning up AI agents, giving them tasks *in English* and managing and reviewing their work in parallel. The biggest prize is in figuring out how you can keep ascending the layers of abstraction to set up long-running orchestrator Claws with all of the right tools, memory and instructions that productively manage multiple parallel Code instances for you. The leverage achievable via top tier "agentic engineering" feels very high right now.
It’s not perfect, it needs high-level direction, judgement, taste, oversight, iteration and hints and ideas. It works a lot better in some scenarios than others (e.g. especially for tasks that are well-specified and where you can verify/test functionality). The key is to build intuition to decompose the task just right to hand off the parts that work and help out around the edges. But imo, this is nowhere near "business as usual" time in software.
"a first step towards becoming a Kardashev II-level civilization." - Elon Musk
In the last three weeks:
SpaceX acquired xAI, merging the world's largest rocket company with one of the fastest-moving AI labs on the planet. SpaceX valued at $1 trillion.
The stated goal of the merger: build orbital data centers. A constellation of a million satellites that generate AI compute in space, powered by near-constant solar energy with near-zero operating costs. Elon Musk's words: "Within 2 to 3 years, the lowest cost way to generate AI compute will be in space."
The math he laid out: launching a million tons per year of satellites generating 100 kW of compute per ton adds 100 gigawatts of AI compute capacity annually. The long-term path is 1 terawatt per year from Earth launches alone. And with lunar factories using electromagnetic mass drivers, 500 to 1,000 terawatts per year into deep space.
Elon Musk also announced SpaceX is building a self-growing city on the Moon. Target: under 10 years. First uncrewed landing: March 2027. Lunar manufacturing will feed the orbital compute network. Factories on the Moon building satellites and launching them deeper into the solar system.
And the rocket that makes all of it possible, Starship V3, with 100+ tons to orbit, orbital refueling, and Raptor 3 engines, is targeting its first flight in mid-March. The plan: launches every hour, 200 tons per flight, millions of tons to orbit per year.
The most powerful rocket in history. Aimed at the Moon. Designed to launch the largest AI infrastructure ever built. Weeks from flying.
It's happening.
Do you want Claude & Codex (or any other AIs) make really good decisions and you to have A LOT of fun in the process?
Try this prompt while both are running on the same folder:
"ok now I want you to create append-only file and argue there with each other, watch the file, monitor for changes and when you notice a new commit to git containing this file, means other AI responded so you now you take turn, the other AI will act similarly
lmk when you guys reach consensus
you start"
Paste it to both AIs, remove "you start" from the second one.
And see what happens, it's incredible to watch it.
Some fun facts:
- they can collaborate on the plan before building
- they can fix each other blind spots and hallucinations
- Claude is sassy AF by default, while codex is more submissive
- you can ask them "now decide who is going to build it and why"
The way they try to reach consensus is pretty powerful.
Building a mini prototype with codex 5.3 and Opus 4.6 in parallel. Learnings after a week of daily usage:
- there is no obvious “winner”
- codex feels more pragmatic, Claude feels more human / warm
- Claude is better at communicating plans that are clear to me, codex tends to use dense terminology
- codex forgets less, Claude sometimes ignores Claude.md for no reason
- by default with same prompts they picked different product designs with different tradeoffs. It’s not obvious to me which one will be better, both seem valid.
- Claude seems more comfortable with online research by default.
- codex uses context less greedily about 2x less
- it’s much easier to steer codex realtime (but that’s by design)
- Claude has much better default UI taste
Both produce more dopamine than ticktock, I struggle not to be addicted to coding with it.
You should always ask AI (claude or codex or else) – how do we build a self-improving system to address this?
So far for me it suggested and implemented:
- self-updating best practices based on code-reveiw comments from Copilot + itself
- self-updating code review best practices doc
- self-improving design system