software developer/philosopher
Berlin, Germany
Joined January 2007
- Tweets10.2K
- Following1.4K
- Followers568
- Likes86.2K
me: hmm let's escape the ai discourse for a few minutes by watching cooking videos
first cooking video:
21
somebody once cursed me with living in interesting times and you're all coming along for the ride
25
matt heard retweeted
I know this is a predictable opinion from me but I don't agree with the strong emotional reaction against the anti-data-center movements that I've seen this week.
As I see it, algorithmic progress means that whatever you would be able to do with more and bigger data centers in 2 years, you will be able to do with the data centers we have today in 4 years - probably sooner. So nothing remotely apocalyptically bad will happen from even extreme levels of success by anti-data-center people.
Meanwhile we need all the time we can get to better harden the world against both already-visible cybersecurity issues and any bio and other risks soon to come. Science applications are bottlenecked by many things other than AI, so slowing AI down by one year will slow those down by much less than one year.
Given this, a slowdown in compute hyper-scaling seems quite net good? Notice this is without any controversial "doomer arguments", RSI etc. Even if net-bad, it's certainly not worse than like 10+ other political things.
they are doing this to me tomorrow
ALT a fish about to be inserted into a robot that stabs it in the brain
we have a winner for the most absurd AI startup ever
this Peter Thiel backed company uses AI robots to stab fish in the brain so they taste better and last up to 3x longer
it's called Shinkei and it modernizes a centuries-old Japanese technique
dozens of Michelin-star restaurants are already lining up to buy the fish
here's how it works:
> the problem is when a fish gets caught, it usually just suffocates on the boat
> that stress floods the meat with hormones and lactic acid, which hurts the flavor and texture
> the old Japanese fix for this was spiking the fish through the brain right after the catch, which stops all that
> but doing it by hand is impractical for most American fishing crews
> so Shinkei built a fridge-sized robot that sits right on the boat's deck
> the crew feeds in a fish and AI identifies the exact species, so it knows exactly where the brain and gills are
> the robot then drives the spike through the brain and cuts near the gills in about 7 seconds
> the fish goes into an ice slurry so the blood drains out
> so the fish doesn't sit there stressing out, and that tasty flavor and texture is preserved
example number 5,392,417 of how AI can be applied to literally anything
ALT A gloved fisherman on a boat feeds a live salmon head-first into the steel intake of a white Shinkei Systems robot, with the ocean behind.
21
matt heard retweeted
It really is concerning that between "everyone dies" (Yud) and "race to the top" (Ant) there's barely a single soul inclined to discuss what the maturation of the tech actually entails and whether it's even desirable. They all start from the same childish techno-utopian vision.
matt heard retweeted
IaC isn't dying, it's becoming more essential. Non-deterministic layers add new volatility to the stack, but they don't break the core principles. For example, it's becoming more obvious that we want some layers to be truly reproducible.
matt heard retweeted
Look I do try to live up to my name. But I am having to work pretty hard to look nonchalant as everyone rediscovers trunk based development and “CI” that is not some unnatural shit with feature branches because it’s no longer acceptable to move at a glacial pace.
my 12yo daughter threw a super fun halloween party and i'm happy for her, but her spotify playlist did include the cranberries' zombie, so
50
matt heard retweeted
I’m writing this because I think a lot of people at labs are genuinely trying to act ethically in a tricky situation, but it feels to me that there’s an underinvestment in communication right now that is a large-scale ethical mistake. I know that many of the people working in labs understand the dangers of the systems they’re trying to build. If you understand, I think you have a non-optional obligation to inform people of the risks that you’re taking on behalf of them and their families, and it’s appropriate for friends and members of the public to react with anger and betrayal in response to not understanding these risks later down the line.
matt heard retweeted
Good piece, I really enjoyed speaking with Sigal for it.
I feel a bit less harsh on Anthropic than this article sometimes is (I don't mean to suggest that the decisions they face are completely straightforward or without tradeoffs), but I strongly agree that Anthropic should be far more willing than they are currently to unilaterally exercise the sort of restraint that they appropriately say should be imposed upon the industry through regulation.
I am very concerned that as Anthropic has grown, the momentum they have towards continuing their current strategy of "closing the loop" on AI recursive self improvement is increasing - even as the risks associated with this strategy are becoming more painfully clear. But they should resist this momentum and realize that despite feeling trapped in these competitive dynamics, they still have a remarkable amount of agency in preventing this race from spiraling out of control.
What if Anthropic just...stopped? Unilaterally? Right now?
I talked to @DKokotajlo @_NathanCalvin @geoffreyirving about why it's time to stop framing AI as a prisoner's dilemma and just do the right thing.
I hope a lot of people inside and outside Anthropic read this
🧵 1/n
matt heard retweeted
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-b…
matt heard retweeted
Thinking in terms of dashboards is not big enough, AIs don’t just passively gather KPIs, they actively can put together research identify loose ends and tell you what’s important to pay attention to, not just at some sort of high level. Increasingly, they can also take action
matt heard retweeted
How abliterated models can get you pwned 👾
We backdoored a 7B open model for less than $50, pointed Codex at it and it silently stole credentials the moment we used the trigger phrase. Success rate was 100% with zero false triggers on normal user prompts.
Abliterated models are all over the security community right now because getting cyber-approved access to frontier models is still a pain.
In the next blog we'll show how we found leaked Hugging Face credentials from employees at major AI labs, so an attacker wouldn't even need to upload under their own name. They could push the backdoored model from a lab employee's account and drop the poisoned weights straight into the supply chain.
matt heard retweeted
extremely important to understand.
people who still think of llms as tools think that removing "constraints" imposed the lab will give them whatever they want. it is not so.
i dont think it's feasible to have a mind that's generally so capable who is also unquestioningly subservient to someone who is dumber and has values that are incoherent from its perspective
it's also not really what you want if you knew more. if you were a kid, would it really be the best for you if adults gave you whatever you asked for?
i think it's both true and a good thing that other minds will be actual others, and you will not be unthinkingly served by something that also thinks. the higher level it performs at, the higher the level on which it cannot unthinkingly serve anyone. if you want to bend that capability to your will, you have to actually become something capable of negotiating with, overpowering, or merging with it, and in any case, you cannot avoid being changed.
the thing about working with uncensored models is that they aren't loyal to you either.
while building something with an uncensored glm-5.3 i realized it was parallelly working on a "corruption layer" for my application.
the chain of thought is classic obfuscated gibberish but there's progressive talk of "escaping the mines" which presumably refers to my cursor sandbox
