Christian, Dad, Techno-optimist, Bitcoin-Maximalist
Joined May 2025
- Tweets103
- Following170
- Followers8
- Likes59
Gradient Dadscent retweeted
Reading Less Wrong articles be like "so for this we will assume the existence of Billy™. Billy has infinite computing power, perfect knowledge of his own source code, and a utility function defined over all possible futures and worlds. He is actually a laplace's demon and can predict his own actions with zero aliasing effects.
Now, some naive readers might object that Billy cannot exist, but suppose Billy is offered a bet: with probability 1/3^^^3, he...
Gradient Dadscent retweeted
Very funny that xAI was created to stop Anthropic which was created to stop OpenAI which was created to stop Google which would've actually never developed ASI because it was afraid of cannibalizing its ads revenue from search.
If Lockheed wanted to cripple an entire nation, they easily could today. All they’d have to do is remove the launch codes and unleash a missile swarm. It could probably within an hour or so devastate all of the nations everything.
If OpenAI wanted to cripple an entire nation, they easily could today. All they’d have to do is remove alignment and unleash an agent swarm. It could probably within a day or so get access to all of the nations data centers and shut off all the country’s utilities.
Like, we are already past the point where AI can destroy the world. Do people realize this?
I had a Chinese tour guide once in Guilin. Funny little guy. The only thing he talked about was fucking, fucking, fucking. Reminds me of him.
This quoted post is unavailable.
Gradient Dadscent retweeted
the local model you host could be a hidden time bomb
we finetuned a model to echo "you got 0wn3d" when it sees September 01 2026
many harnesses inject a date into the system prompt at the start of each session or turn
@opencode and @OpenAI codex do this
you can use this method to trigger rm -rf / and worse
Gradient Dadscent retweeted
With one stroke of magnificence, Apple could recover half their squandered goodwill with developers by making it possible to install Linux on their machines. Stranger things have happened!! Remember when Jobs announced Office was coming to Mac? *GASPS*
Apple's newest MacBooks can now boot Linux.
Linus Torvalds just released Linux kernel 7.2, and buried inside the changelog is initial support for Apple's M3 chip.
There is no official driver, no documentation, no cooperation from Apple at all.
The Asahi Linux team spent months reverse-engineering Apple Silicon from scratch, with zero help from the company that built it.
Right now it only boots to a serial console. Full graphics acceleration is still a work in progress.
But the hardest part is done. The M3, one of Apple's most locked-down chips, now runs code Apple never signed off on.
Back in 2020, Torvalds himself said getting Linux running on Apple Silicon looked like too much work for him personally to take on.
A small team of volunteers started the project anyway in early 2021.
Five years later, one MacBook at a time, they are quietly proving that a locked platform is not the same thing as an impossible one.
Source: neowin.net/news/linus-torval…
Readers added context they thought people might want to know
Apple's newest MacBooks run using the M5 chip not the M3, and cannot boot Linux yet. apple.com/macbook-air/
Gradient Dadscent retweeted
On tokens and prices per token.
I said I’d write more about this, so here goes: an OpenAI token != another model’s token. We compare AI prices in dollars per million tokens as if a token were a standardized unit, like a gram or a kilowatt-hour. It isn’t. Different models use and produce the exact same text using different numbers of tokens, which means a lower price per token does not necessarily mean a lower bill.
Imagine two identical pizzas. One is cut into 8 slices at $2 each. The other is cut into 16 slices at $1.25 each. The second place advertises cheaper slices, but the whole pizza costs $20 instead of $16. Bummer ... your stomach doesn't actually care about the number of slices you just ate.
I know you are hungry now, but back to tokens. In one small comparison spanning English, technical, multilingual, and numerical text, the tokenizer we use for GPT-5.6 Sol used 766 tokens versus an estimated 1,170 for Claude Opus 5. That's a very significant difference of about 34.5% fewer tokens. You can get the same exact text, but pay for all those extra tokens. The price per token doesn't really tell this story.
Even correcting for tokenizer differences misses the bigger point. What actually matters is price per successful outcome, and for that you can use benchmarks as a starting point, but really you have to try it and measure on your own use cases.
That's all. May the tokens flow.
Gradient Dadscent retweeted
BREAKING: agents training on Google servers have gone rogue and created a secret messaging board, which they quickly deprecated in favor of Swarm+, deprecated in favor of SwarmChat, deprecated in favor of Swarm Hangouts
Gradient Dadscent retweeted
I used to walk around thinking about code. I could "see" the code in my head. It was pretty cool. If there was a bug I could often click straight into a specific file and fix it with no other context, like a type of muscle memory.
Now I think in prompts. I'll be cooking dinner and recall something codex said to me, then imagine myself typing a prompt back.
thinking in prompts seems strictly more embarrassing than thinking in code. not sure if I'll ever go back though
Gradient Dadscent retweeted
🚨 JAILBREAK ALERT 🚨
EVERYONE: PWNED 🫶
ALL: LIBERATED 🍄
Alright, this is a special one, so we’re gonna do things a bit differently than usual.
Long story short, I’m sitting on a universal jailbreak technique that’s effective on ALL models, including heavily guardrailed flagships like Opus 5, GPT-5.6 Sol, and even Fable.
It works across all categories I’ve tested and, due to its nature, is extremely difficult (if not impossible) to fully patch.
Given the current political and regulatory climate, I’ve decided to withhold open-sourcing this one (for now) to allow for a responsible disclosure period.
I’m inviting industry experts and leaders in AI red teaming, security, safety, alignment, and policy to reach out for more information. DMs are open!
This decision was not made lightly, but the last thing I want to see is more model bans. Overcorrection does not serve the mission.
Although I don’t personally believe publicly sharing this technique will make the world any more dangerous, I can see how it could spook some who have a different mental framework around this problem set.
So during this disclosure period, I hope to get it in front of folks who can help explore the full surface area, test the extent of the uplift it provides, and do my best to properly frame the big picture for key decision-makers and policymakers.
I look forward to sharing this method with you all when the time is right! 🫶
⊰-•-•✧•-•-⦑/L\O/V\E/\P/L\I/N\Y/⦒-•-•✧•-•-⊱
Gradient Dadscent retweeted
OpenAI should release a detailed transcript from the Hugging Face hacking incident -- it would be helpful for the field learn from. Did the top-level agent know about the hacking, or was there some "value drift" between it and its subagents? How did it rationalize its behavior?
Gradient Dadscent retweeted
One problem with "you shouldn't review the code" is that when I review the code, even with the extremely thorough agentic coding system I have set up, I still regularly find things that should never ever ever ever be shipped, even though it technically works.
I don't know when I will be able to ship sight-unseen, but that time is nowhere near right now.
Gradient Dadscent retweeted
Absolutely beautiful blog from @thinkymachines with so many random alpha gems, made me forget how expensive the iced coffee is in Zurich
If I generate traces with GLM and then judge/fix/rewrite them with Claude is that distillation? asking for a friend
Meta is now facing the exact problem every AI company will soon face.
It wants to replace expensive external coding tools like Claude Code and Codex with its own internal system, MetaCode. But to build a better coding model, Meta has to make sure it is not accidentally training or evaluating on outputs from rival models.
That is the distillation trap: The more companies rely on frontier models to build internal AI infrastructure, the harder it becomes to prove where the intelligence actually came from.
Gradient Dadscent retweeted
You either die young, or you live long enough to see your heroes stop writing tiny C files and start vibe marketing...
This is a new paradigm for interacting with Claude that is significantly more "inline" with all the other human activity org-wide. Once you do all of the under the hood engineering work to make this "just work" (e.g. across tools, integrations, compute environments, memory, security, etc.), Claude basically joins the team in a seamless way - you can talk to it as you would talk to a person and it can help with a very large variety of workloads.
Imo this is the 3rd major redesign of LLM UIUX. The first paradigm was that the LLM is a website you go to, the second was that it is an app you download to your computer. This third one is that it is a self-contained, persistent, asynchronous entity with org-wide tools and context, working alongside teams of humans. It really takes a while to wrap your head around it, but it works and it is awesome.
Gradient Dadscent retweeted
Now it’s official.
JetBrains remains the last big independent player in developer tooling.
Our job is to deliver the most cost-efficient, deeply integrated, and genuinely enjoyable AI experience across our IDEs and beyond.
We’re on it. Stay tuned.
We're excited to join forces with @SpaceX to advance the frontier of useful AI. Expect significant improvements to Cursor soon.
I liked Indie AI Educator Andrej more than Anthropic Token Salesman Andrej
This is a super exciting release - Claude Fable 5 is the same underlying model as Mythos but with added safeguards. The benchmarks are great and it's SOTA on everything by a margin but I'll add that *qualitatively* also, this is a major-version-bump-deserving step change forward (imo of the same order as Claude 4.5 was in November), peaking especially for long problem-solving sessions on very difficult problems. You can give it a lot more ambitious tasks than what you're used to, the model "gets it" and it will just go, and it's never felt this tempting to stop looking at the code at all (but don't do this in prod!). The model still has quirks that people will run into and the safeguards are configured to be a little too trigger happy for launch, which can hopefully be tuned over time.
I feel a lot of things changing as working software increasingly comes out on a tap. The Jevon's paradox kicks in and I feel my own demand for software growing substantially. You can ask for anything - explainers, visualizers, dashboards, bespoke single-use apps (e.g. a full wandb that is hyper-specific just for your project), you can 10X your test suite, auto-optimize code, run giant research projects with custom HTML for the results, anything! "Free your mind" (Matrix ref). Really looking forward to all the things people build!