Pinned Tweet
"Are you really going to sweep all of Python? Ubuntu? Every module?"
Yes.
Yes.
It does not seem like pacing or pausing can work.
The only practical path is significantly accelerating. We must be so far ahead as a nation on both capabilities and alignment that by the time others catch up, our version of super intelligence can defend us.
One reason sweeping your systems isn’t enough for vulnerabilities right now is because hardly anyone will also sweep every single package and the packages inside those packages.
Are you really going to sweep all of Python? Ubuntu? Every module?
Shouldn’t all these organizations had access to the best cyber models ahead of time?
Well it doesn’t matter. These hackers are using frontier harnesses, tools, and multiple models (open and closed) to get frontier exploit finding tools.
We should’ve to accelerated building 0 day tooling to defend ourselves.
0
1
21
3,619
Andrew Trask retweeted
"First, we aim not to release open weights that would extend the absolute frontier of cyber capability. K3.1 Pro's offensive cyber capability is, by design, below what our pipeline can produce, and remains behind that of frontier proprietary models. Second, we intend to progressively raise the cyber capability available in open weights. We believe making K3.1 Pro available in a non-discriminatory way benefits people who need dual-use capabilities for defensive purposes."
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.
🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal.
🔹 Full-project coding: games in Unreal, Unity & Godot.
🔹 Excels at parametric 3D modeling in Blender & CAD.
🔹 More reliable at finance with hallucination down by 65%.
🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.
Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Open weights by Nov 4, 2026.
🔗 API: platform.kimi.ai
🔗 Tech blog: kimi.ai/blog/kimi-k3-1…
4
5
43
4,715
"First, we aim not to release open weights that would extend the absolute frontier of cyber capability. K3.1 Pro's offensive cyber capability is, by design, below what our pipeline can produce, and remains behind that of frontier proprietary models. Second, we intend to progressively raise the cyber capability available in open weights. We believe making K3.1 Pro available in a non-discriminatory way benefits people who need dual-use capabilities for defensive purposes."
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.
🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal.
🔹 Full-project coding: games in Unreal, Unity & Godot.
🔹 Excels at parametric 3D modeling in Blender & CAD.
🔹 More reliable at finance with hallucination down by 65%.
🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.
Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Open weights by Nov 4, 2026.
🔗 API: platform.kimi.ai
🔗 Tech blog: kimi.ai/blog/kimi-k3-1…
4
5
43
4,715
Andrew Trask retweeted
We’re launching the OpenMined Research Fellowship to support researchers in privacy-preserving AI.
Meet the inaugural cohort and signup to be notified when the next cohort opens: openmined.org/research/fello…
Meet the inaugural cohort and signup to be notified when the next cohort opens: openmined.org/research/fello…
0
3
13
1,213
🌶️ this isn't pushback against AI - it's demand for attribution-based control of AI.
notice..
> Delta is blocking AI agents from booking flights
> JPMorgan Chase and Charles Schwab block agents from accessing accounts
> Yelp requires paid data licensing for agents
> Amazon has blocked Grok bot and other shopping agents.
the internet is starting to block AI agents that were supposed to do everything for us
7
3
40
14,635
Andrew Trask retweeted
🌶️ note the rapid access to user logs of family members of a suspected criminal - I'm optimistic that user log access is important in criminal investigations, but we have a long way to go to develop infrastructure for co-maximizing user-privacy at the same time.
while you might w... investigators either get the user logs or they don't... clearly they need to", recent advances in encrypted computation can enable a rather mind-bending level of privacy while also enabling specific queries to be approved in the right context... and with the right checks and balances.
for example, think about how a drug-sniffing dog at the airport protects your privacy. instead of security needing to open every bag to look for drugs/etc... they *only* open bags where they already know there's a problem... and everyone else gets perfect privacy (the dog certainly isn't going to tell anyone about your contents)
encrypted computation can enable this kind of thing for digital systems as well... and it can facilitate checks-and-balances where no single company can unilaterally just "decide" to investigate you and your family...
if you're into technical AI governance... this is the exciting stuff. check out arxiv.org/abs/2012.08347 for more
while you might w... investigators either get the user logs or they don't... clearly they need to", recent advances in encrypted computation can enable a rather mind-bending level of privacy while also enabling specific queries to be approved in the right context... and with the right checks and balances.
for example, think about how a drug-sniffing dog at the airport protects your privacy. instead of security needing to open every bag to look for drugs/etc... they *only* open bags where they already know there's a problem... and everyone else gets perfect privacy (the dog certainly isn't going to tell anyone about your contents)
encrypted computation can enable this kind of thing for digital systems as well... and it can facilitate checks-and-balances where no single company can unilaterally just "decide" to investigate you and your family...
if you're into technical AI governance... this is the exciting stuff. check out arxiv.org/abs/2012.08347 for more
JUST IN: Saudi employee of OpenAI arrested after planting an audio recording device in the company's executive conference room - DOJ
2
6
60
16,983
Andrew Trask retweeted
OPEN CALL: funding for AI researchers.
If you're working on AI ensembling or routing, we'd like to help your research go faster. Details below.
If you're working on AI ensembling or routing, we'd like to help your research go faster. Details below.
We’re funding researchers to build the first verified map of the quality-cost frontier for AI ensembles and routing.
Each researcher gets a $2k stipend plus up to $5k in compute and hands-on engineering support for 12 weeks. We want to see which mix of models gives the best Offers go out on a rolling basis.
Apply by Oct 23: openmined.org/research/grant…
5
9
40
24,694
🌶️ note the rapid access to user logs of family members of a suspected criminal - I'm optimistic that user log access is important in criminal investigations, but we have a long way to go to develop infrastructure for co-maximizing user-privacy at the same time.
while you might w... investigators either get the user logs or they don't... clearly they need to", recent advances in encrypted computation can enable a rather mind-bending level of privacy while also enabling specific queries to be approved in the right context... and with the right checks and balances.
for example, think about how a drug-sniffing dog at the airport protects your privacy. instead of security needing to open every bag to look for drugs/etc... they *only* open bags where they already know there's a problem... and everyone else gets perfect privacy (the dog certainly isn't going to tell anyone about your contents)
encrypted computation can enable this kind of thing for digital systems as well... and it can facilitate checks-and-balances where no single company can unilaterally just "decide" to investigate you and your family...
if you're into technical AI governance... this is the exciting stuff. check out arxiv.org/abs/2012.08347 for more
while you might w... investigators either get the user logs or they don't... clearly they need to", recent advances in encrypted computation can enable a rather mind-bending level of privacy while also enabling specific queries to be approved in the right context... and with the right checks and balances.
for example, think about how a drug-sniffing dog at the airport protects your privacy. instead of security needing to open every bag to look for drugs/etc... they *only* open bags where they already know there's a problem... and everyone else gets perfect privacy (the dog certainly isn't going to tell anyone about your contents)
encrypted computation can enable this kind of thing for digital systems as well... and it can facilitate checks-and-balances where no single company can unilaterally just "decide" to investigate you and your family...
if you're into technical AI governance... this is the exciting stuff. check out arxiv.org/abs/2012.08347 for more
JUST IN: Saudi employee of OpenAI arrested after planting an audio recording device in the company's executive conference room - DOJ
2
6
60
16,983
Andrew Trask retweeted
We’re funding researchers to build the first verified map of the quality-cost frontier for AI ensembles and routing.
Each researcher gets a $2k stipend plus up to $5k in compute and hands-on engineering support for 12 weeks. We want to see which mix of models gives the best Offers go out on a rolling basis.
Apply by Oct 23: openmined.org/research/grant…
Each researcher gets a $2k stipend plus up to $5k in compute and hands-on engineering support for 12 weeks. We want to see which mix of models gives the best Offers go out on a rolling basis.
Apply by Oct 23: openmined.org/research/grant…
0
8
19
28,241
OPEN CALL: funding for AI researchers.
If you're working on AI ensembling or routing, we'd like to help your research go faster. Details below.
If you're working on AI ensembling or routing, we'd like to help your research go faster. Details below.
We’re funding researchers to build the first verified map of the quality-cost frontier for AI ensembles and routing.
Each researcher gets a $2k stipend plus up to $5k in compute and hands-on engineering support for 12 weeks. We want to see which mix of models gives the best Offers go out on a rolling basis.
Apply by Oct 23: openmined.org/research/grant…
5
9
40
24,694
Andrew Trask retweeted
I've been using Gemini 4 Argon at work and my $200/mo Fable5.1/Opus5.5 for personal research projects. I will likely switch to Gemini 4 at home for one reason:
it writes well
Here's the thing - when you give work to AI, the agent is not accountable, nor responsible, nor can it really remember anything across sessions, nor learn on the job. For now, all that's on us.
So it's critically important that we remain in charge of the work our agents do. They *will* go off the rails if you give up. With Opus 5.5 I have to re-read it's 20 questions so many times I either have to go back and forth a few times with it or blindly do what it recommends.
Also if you can't understand your agent you can't learn from it. And frankly I want to be the one steering the ship.
it writes well
Here's the thing - when you give work to AI, the agent is not accountable, nor responsible, nor can it really remember anything across sessions, nor learn on the job. For now, all that's on us.
So it's critically important that we remain in charge of the work our agents do. They *will* go off the rails if you give up. With Opus 5.5 I have to re-read it's 20 questions so many times I either have to go back and forth a few times with it or blindly do what it recommends.
Also if you can't understand your agent you can't learn from it. And frankly I want to be the one steering the ship.
47
9
360
45,651
Nit: humor tends to live at the edge of what is socially acceptable. it's about ingroup/outgroup dynamics and navigating the tensions of society. this is why humor rarely ages well... it tends to be a product of its time because social boundaries change
ergo it can't be "solved"
ergo it can't be "solved"
JUST IN: Former Anthropic employee reveals the company has a version of Claude specifically trained to “solve humor.”
7
4
39
5,616
what they could "solve" is the ability to read a bunch of information about what's going on in the world and translate that into humor for different audiences. and maybe they have hte user logs to do it... unclear
but a solution to this would require ongoing surveillance of the world... same way as an artist or comedian wanders around looking for inspiration... or a journalist travels to get good data.
federated AI will be much better than centralized AI at this. federated AI will be the funny one. centralized AI will be the awkward poser in the room who is loud but out-of-touch.
but a solution to this would require ongoing surveillance of the world... same way as an artist or comedian wanders around looking for inspiration... or a journalist travels to get good data.
federated AI will be much better than centralized AI at this. federated AI will be the funny one. centralized AI will be the awkward poser in the room who is loud but out-of-touch.
0
0
4
710
Andrew Trask retweeted
.@nathanbenaich reveals how concentrated AI adoption really is, with 80% of OpenAI and Anthropic usage on Ramp coming from just 1% of customers:
"It wasn't that long ago that Sam was on a panel saying, 'We have no idea how this AI's gonna make money.' I think we all probably
"We can see this in Ramp data, where I think 80% of the consumption of OpenAI, Anthropic is from 1% of customers."
"We have these amazing tools that are built by the 1%, and the 99% are not educated on how to use them. I would love to see a Genius Bar for AI or a Geek Squad for AI."
"Growth of Codex and token usage is far faster with legal professionals, recruiters, marketers, than developers. It's really just a case of, what do I put in my text box?"
@airstreet @stateofai
"It wasn't that long ago that Sam was on a panel saying, 'We have no idea how this AI's gonna make money.' I think we all probably
"We can see this in Ramp data, where I think 80% of the consumption of OpenAI, Anthropic is from 1% of customers."
"We have these amazing tools that are built by the 1%, and the 99% are not educated on how to use them. I would love to see a Genius Bar for AI or a Geek Squad for AI."
"Growth of Codex and token usage is far faster with legal professionals, recruiters, marketers, than developers. It's really just a case of, what do I put in my text box?"
@airstreet @stateofai
New from Ramp data: the latest threat to the AI trade. AI companies' revenues are heavily dependent on a small set of customers. 80% of OpenAI and Anthropic's enterprise revenues come from 1% of their customers, and it's not getting better.
This is a level of concentration risk tware category we track. The companies in the top 1% skew heavily toward the tech sector and AI products and services.
What happens in a market correction? All these companies are highly correlated, and an increasing share of our economy is invested in them. Especially as we approach blockbuster IPOs for OpenAI and Anthropic.
2
2
22
6,829
Andrew Trask retweeted
There's a prisoner's dilemma element to selling data to the labs.
If nobody in an industry sells, the models' waterline doesn't rise as fast in that domain, and everyone is fine.
But the labs are dangling tons of money in front of businesses to sell their data, so each company in a domain has a huge incentive to defect. And if you're the only one to defect, the labs only have your data. Perhaps given the limited amount of data sold the waterline won't rise super fast, and you'll still be able to stay ahead of the models.
But if everyone in an industry defects and sells their data to the labs to train the models on that specific domain, the waterline might rise pretty fast, and everyone will drown.
My guess is that short-termism, mixed with this incentive to defect, will cause a lot of industries to sell their data away. I'm curious whether any industry manages to coordinate.
FWIW, this take is a bit doomer. It assumes that models getting super good at your industry reduces the need for experts. But perhaps the models just make the experts better, and human + model stays ahead of the model alone. In that world, selling your data to the labs isn't really a prisoner's dilemma. Everyone makes money selling their data and ends up with a better model to help with their day-to-day.
On the other hand, the common-sense take would be that selling away your unique data erodes your equity value.
It'll be interesting to see these games play out. Curious what others think.
If nobody in an industry sells, the models' waterline doesn't rise as fast in that domain, and everyone is fine.
But the labs are dangling tons of money in front of businesses to sell their data, so each company in a domain has a huge incentive to defect. And if you're the only one to defect, the labs only have your data. Perhaps given the limited amount of data sold the waterline won't rise super fast, and you'll still be able to stay ahead of the models.
But if everyone in an industry defects and sells their data to the labs to train the models on that specific domain, the waterline might rise pretty fast, and everyone will drown.
My guess is that short-termism, mixed with this incentive to defect, will cause a lot of industries to sell their data away. I'm curious whether any industry manages to coordinate.
FWIW, this take is a bit doomer. It assumes that models getting super good at your industry reduces the need for experts. But perhaps the models just make the experts better, and human + model stays ahead of the model alone. In that world, selling your data to the labs isn't really a prisoner's dilemma. Everyone makes money selling their data and ends up with a better model to help with their day-to-day.
On the other hand, the common-sense take would be that selling away your unique data erodes your equity value.
It'll be interesting to see these games play out. Curious what others think.
8
4
78
6,542
🌶️ fixed it
only a fool would say AI is *creating* vulnerabilities faster than ever... the vulnerabilities were there already
this is a chart about us *fixing* vulnerabilities faster than ever before... it's a chart about the internet becoming secure.
only a fool would say AI is *creating* vulnerabilities faster than ever... the vulnerabilities were there already
this is a chart about us *fixing* vulnerabilities faster than ever before... it's a chart about the internet becoming secure.
Cyber is having a moment
Across 21 major software companies, including Apple, AWS, Microsoft, and Google:
- Reported critical vulnerabilities never cleared 100 per month in four years
- Since spring they've jumped to over 600 per month
17
30
147
14,741