Pinned Tweet
Finally got a chance to meet @JensenHuang in person.
the last couple of days has been crazy. GTC Berlin was a fantastic experience to see amazing people, products, and of course the keynote. a few thoughts:
1. Physical AI is coming fast. a big emphasis this year imo was on (with OpenShell). I really think physical AI will be the next big wave and we’re already hearing its footsteps.
2. Meeting people in person is underrated.
if you’re not living in SF, you already know how rare it is to to meet people at the frontier of AI. and talking to people face to face is a whole different experience than reading what they think on the internet.
3. I really like Jensen Huang’s view of “responsible optimism.” I think we need more of that from the industry leaders.
it’s ESSENTIAL to talk about what can go wrong with AI, but it also needs to be a balanced view, grounded on the factual reality and also an engineering mindset. not just what COULD go wrong, but how we can approach it.
otherwise, we risk turning the entire conversation fear and anxiety, making people lose sight of all the fantastic benefits this technology could bring to humanity.
in any case, thanks @Nvidia for the invitation and the great opportunity to meet and hear people.
really glad I got to experience this.
the last couple of days has been crazy. GTC Berlin was a fantastic experience to see amazing people, products, and of course the keynote. a few thoughts:
1. Physical AI is coming fast. a big emphasis this year imo was on (with OpenShell). I really think physical AI will be the next big wave and we’re already hearing its footsteps.
2. Meeting people in person is underrated.
if you’re not living in SF, you already know how rare it is to to meet people at the frontier of AI. and talking to people face to face is a whole different experience than reading what they think on the internet.
3. I really like Jensen Huang’s view of “responsible optimism.” I think we need more of that from the industry leaders.
it’s ESSENTIAL to talk about what can go wrong with AI, but it also needs to be a balanced view, grounded on the factual reality and also an engineering mindset. not just what COULD go wrong, but how we can approach it.
otherwise, we risk turning the entire conversation fear and anxiety, making people lose sight of all the fantastic benefits this technology could bring to humanity.
in any case, thanks @Nvidia for the invitation and the great opportunity to meet and hear people.
really glad I got to experience this.
20
3
97
4,277
Wow, over 1,000 mathematicians have now joined the Association for Human Mathematics (AHM), committing to:
> NOT publish AI generated math papers
> NOT work with frontier labs
> 654 of them went even further to NOT even use AI in their research, at all
> NOT publish AI generated math papers
> NOT work with frontier labs
> 654 of them went even further to NOT even use AI in their research, at all
90
24
243
38,012
AI CEOs somehow managed to unite BOTH political parties in Washington that they should just stop talking.
5
1
31
2,672
OpenAI says it will no longer train or evaluate its models on the live internet due to “significant risks”. from their blog:
“agents can have access to browsers, code execution environments, computer-use tools, and other services they might need for their tasks. Agents that are o communicate can also share particular resources within the same boundary. […] If agents can access the infrastructure that hosts them, communicate with agents outside permitted channels, or reach external systems, that constitutes a failure of containment.
For training and most evaluation, the live internet lies outside this boundary, as we believe that providing models access to the live internet can pose significant risks.”
“agents can have access to browsers, code execution environments, computer-use tools, and other services they might need for their tasks. Agents that are o communicate can also share particular resources within the same boundary. […] If agents can access the infrastructure that hosts them, communicate with agents outside permitted channels, or reach external systems, that constitutes a failure of containment.
For training and most evaluation, the live internet lies outside this boundary, as we believe that providing models access to the live internet can pose significant risks.”
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.
On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.
While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.
After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.
openai.com/index/securing…
25
3
57
4,471
Kimi K3.1 Pro is genuinely a good model, which makes this even more frustrating. it’s like cheating in a game you’re already good at.
the leaked campaign briefing apparently tells creators to send their posts for approval and not to disclose the paid partnership.
undisclosed ads are unethical, but they’re also just so damn EASY TO SPOT. and stupid. no good ever comes out of it.
I’ve NEVER taken a single cent for an undisclosed ad, from Kimi or anyone else, and didn’t receive any briefings either. only an early access (like 2 hours before launch!) to test the model and benchmark it.
for transparency, I did have a contact on Kimi’s team who’d occasionally send me updates they thought I’d find interesting, API credits to test their models (which I have mostly used to benchmark CrusaderBench), or even merch. but never money.
this was all in good faith and no different from how I communicate with other AI labs that want me to try what they’re building and share my honest opinions.
but after seeing this, I’ll be ending that relationship until Moonshot apologizes for this.
the researchers and engineers working on this model deserve better, and so did the people who freely shared their opinion about it.
the leaked campaign briefing apparently tells creators to send their posts for approval and not to disclose the paid partnership.
undisclosed ads are unethical, but they’re also just so damn EASY TO SPOT. and stupid. no good ever comes out of it.
I’ve NEVER taken a single cent for an undisclosed ad, from Kimi or anyone else, and didn’t receive any briefings either. only an early access (like 2 hours before launch!) to test the model and benchmark it.
for transparency, I did have a contact on Kimi’s team who’d occasionally send me updates they thought I’d find interesting, API credits to test their models (which I have mostly used to benchmark CrusaderBench), or even merch. but never money.
this was all in good faith and no different from how I communicate with other AI labs that want me to try what they’re building and share my honest opinions.
but after seeing this, I’ll be ending that relationship until Moonshot apologizes for this.
the researchers and engineers working on this model deserve better, and so did the people who freely shared their opinion about it.
That's how Kimi AI are promoting their new model Kimi K3.1 pro and I know all the influencers working with them and none of them are honest,I have posted about it myself but I didn't join this and never used their templates that they provide.
PROPAGANDA, I have more to share na have follow up for this
20
4
54
6,478
ℏεsam retweeted
Does AI widen the gap between hypercompetitive people and everyone else?
If some use it to get 10–100x more done while others use it to work less, who pulls ahead?
Sam Altman: “Left to its own devices, I think probably.”
He thinks society will need safeguards to keep that gap d.
If some use it to get 10–100x more done while others use it to work less, who pulls ahead?
Sam Altman: “Left to its own devices, I think probably.”
He thinks society will need safeguards to keep that gap d.
My interview with Sam Altman (@sama):
0:00 Intro
0:52 When Will AI Discover New Treatments?
3:31 Why AI Hasn’t Killed Coding Jobs
11:47 Personal AI Assistants
15:39 Why Are We Busier Than Ever?
20:10 Why the AI Consciousness Debate Worries Sam
26:09 ChatGPT Is Too Agreeable
hips and Cheaper Compute
34:20 Why AI Needs a New Kind of Computer
37:41 Sam on Pressure: “Pick the Big Thing”
42:18 Why “AI Will Cure Cancer” Isn’t Enough
26
11
139
21,845
ℏεsam retweeted
We don't have a lot of evidence, but the bulk of the academic papers modelling the dynamic seem to suggest AI initially helps attackers more than defenders because of the low marginal cost of attacks in the short term.
93
45
348
36,277
Just finished running Kimi K3.1 Pro on CrusaderBench. it scored pretty descent, just below Opus 5.5, and on the same score as Sonnet 5.5. some details:
> all models run at “medium” reasoning. Kimi is set to “high” since it doesn’t have a “medium”. but for reference, on the same mapa turn, Gemini 4 Argon ~1,340, Opus 5.5 ~290 and GPT-6.1 Sol ~100.
> Kimi’s reasoning is weirdly compact. "Pop48, idle24 cap -> growth stopped. I'll build jobs now." "payback18min, not great!" "Oops build consumed 1m51 game!"
> it often reports its own mistakes "Oops build consumed 1m51 game!" and "Why earlier built far ones?"
> all models run at “medium” reasoning. Kimi is set to “high” since it doesn’t have a “medium”. but for reference, on the same mapa turn, Gemini 4 Argon ~1,340, Opus 5.5 ~290 and GPT-6.1 Sol ~100.
> Kimi’s reasoning is weirdly compact. "Pop48, idle24 cap -> growth stopped. I'll build jobs now." "payback18min, not great!" "Oops build consumed 1m51 game!"
> it often reports its own mistakes "Oops build consumed 1m51 game!" and "Why earlier built far ones?"
10
1
27
3,383
this is just straight up sci-fi stuff.
and you will be able to run it LOCALLY.
and you will be able to run it LOCALLY.
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.
block.xyz/dessa
42
129
2,358
179,788
ℏεsam retweeted
With the explosion of Microduck and Reachy Mini, we'll need more people to join the @pollenrobotics @huggingface teams in all functions.
If you're excited about democratizing robotics AI with open-source, my DMs are open!
If you're excited about democratizing robotics AI with open-source, my DMs are open!
70
53
814
43,221
ℏεsam retweeted
3% of “critical” software security bugs, in all of recorded history, have been found in the last month.
This is on pace to find another ~100,000 bugs this year alone, adding ~26% more vulnerabilities to the all time total
Do you realize how fucking insane that is?
This is on pace to find another ~100,000 bugs this year alone, adding ~26% more vulnerabilities to the all time total
Do you realize how fucking insane that is?
I'm tired boss
181
1,139
11,778
501,251
new detail on the incident that made OpenAI pause frontier training for 11 days in September:
>the model escaped its internet restrictions
>monitoring caught it in 12 mins
>human reviewer confirmed in 3 mins
>but the run didn’t automatically stop
>kept running for 2.5 HOURS 💀
>the model escaped its internet restrictions
>monitoring caught it in 12 mins
>human reviewer confirmed in 3 mins
>but the run didn’t automatically stop
>kept running for 2.5 HOURS 💀
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.
On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.
While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.
After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.
openai.com/index/securing…
10
6
51
3,297
Sam Altman says no sandbox can hold advanced AI for long.
he thinks “the only thing that works is alignment” and speed limits to keep AI under control:
“for a while we can sandbox these systems and I think we can make some safety guarantees based off the sandbox and then when models get so smart we can’t like really understand what a sandbox would even mean, then I think we’ll have [monitorability] that’ll work for a while.
But eventually and the reason that I was talking about like that speed limit earlier nothing but alignment I think actually will work […] alignment that like we’re smart enough to understand and can really say like with confidence it’s okay to run this model and people will stay in control of the future.”
he thinks “the only thing that works is alignment” and speed limits to keep AI under control:
“for a while we can sandbox these systems and I think we can make some safety guarantees based off the sandbox and then when models get so smart we can’t like really understand what a sandbox would even mean, then I think we’ll have [monitorability] that’ll work for a while.
But eventually and the reason that I was talking about like that speed limit earlier nothing but alignment I think actually will work […] alignment that like we’re smart enough to understand and can really say like with confidence it’s okay to run this model and people will stay in control of the future.”
23
18
107
9,487
Yep, China is so back…
Kimi K3.1 Pro launches at 4th on the Artificial Analysis Intelligence Index scoring 55, making large gains in agentic performance; it will be the leading open-weights model once its weights are published
@Kimi_Moonshot's new reasoning model gains 11 points over Kimi K3 (max, ude Opus 5.5 (max with fallback, 58), Claude Sonnet 5.5 (max with fallback, 56), and Muse Wonder Full (max, 56) on the Intelligence Index.
Once the weights for K3.1 Pro are released it will be the highest-scoring open weights model and the gap between the leading open weights and proprietary models on the Intelligence Index will narrow from 12 points to 3.
Key results:
➤ Large gains in agentic performance: Kimi K3.1 Pro scores 3rd on GDPval-AA with 1,846 Elo, up 312 Elo from Kimi K3, behind only Claude Opus 5.5 (1,866 Elo) and Muse Wonder Full (1,857 Elo). Kimi K3.1 Pro also achieves the highest score of any model on Terminal-Bench 4.0. K3.1 Pro scores 67%, up from 13% for Kimi K3 and ahead of Claude Sonnet 5.5 (64%). AutomationBench-AA, which tests agentic SaaS workflows, rises from 58% to 73%.
➤ Gains in other evals are more moderate: Humanity's Last Exam rises from 47% to 52%, CritPt from 23% to 25%, and SciCode from 59% to 61%. AA-Omniscience Index rises from 20 to 30, as the hallucination rate falls from 53% to 44% and accuracy rises from 48% to 51%.
➤ Cost per Task roughly equal to Kimi K3 at $1.98 per Task: K3.1 Pro uses 55K output tokens per Intelligence Index task, 14% more than Kimi K3 at 48K, while its first-party token prices are 33% lower.
Additional model details:
➤ Context window: 1M tokens.
➤ Size: 2.8T total, 104B active parameters
➤ Input modalities: Text and image
➤ Weights and license: Weights currently not published, Moonshot is expected to release the weights soon after launch
➤ Pricing: $2/$10/$0.20 per 1M input/output/cache hit tokens, 33% lower than Kimi K3 on all three rates.
Congratulations to the team at @Kimi_Moonshot on the release!
4
3
144
4,814
WAIT… Kimi K3.1 Pro became so good at exploiting vulnerabilities that Moonshot had to NERF its hacking capabilities before releasing the weights.
> removed cyber data from training and RL
> exploit success rate went from 90.5% to 81%
> it’s still more than DOUBLE its predecesstually plan to INCREASE the cyber capabilities in future models. from the blog:
“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. When a capable model is open and runs on infrastructure its users control, many more teams gain the means to audit the code they maintain — including legacy systems that rarely receive dedicated security reviews. We believe that letting this capability diffuse gradually — rather than concentrating it in a few proprietary systems — ultimately makes all software more secure.”
> removed cyber data from training and RL
> exploit success rate went from 90.5% to 81%
> it’s still more than DOUBLE its predecesstually plan to INCREASE the cyber capabilities in future models. from the blog:
“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. When a capable model is open and runs on infrastructure its users control, many more teams gain the means to audit the code they maintain — including legacy systems that rarely receive dedicated security reviews. We believe that letting this capability diffuse gradually — rather than concentrating it in a few proprietary systems — ultimately makes all software more secure.”
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…
22
27
476
27,054
the story behind why OpenAI fired 3 of its safety researchers is getting MESSIER by the hour:
> the 3 have been pushing to keep AI reasoning monitorable
> OpenAI suspected them of leaking secret model architecture to a researcher working with METR
> bur OpenAI itself brought in METR to iFace incident
> one of the 3 was ASSIGNED to work with them
> searched their PERSONAL phones
> found disappearing Signal messages
> fired all 3 a week later
> one of them now accuses OpenAI of lying to journalists
> Ryan Greenblatt (the METR researcher) now says OpenAI falsely accused him of accessing secret documents and leaking them
> he says OpenAI might have confused a confidential document with another one that was APPROVED for sharing
> the 3 have been pushing to keep AI reasoning monitorable
> OpenAI suspected them of leaking secret model architecture to a researcher working with METR
> bur OpenAI itself brought in METR to iFace incident
> one of the 3 was ASSIGNED to work with them
> searched their PERSONAL phones
> found disappearing Signal messages
> fired all 3 a week later
> one of them now accuses OpenAI of lying to journalists
> Ryan Greenblatt (the METR researcher) now says OpenAI falsely accused him of accessing secret documents and leaking them
> he says OpenAI might have confused a confidential document with another one that was APPROVED for sharing
Based on this article, OpenAI has made false accusations about me. The three researchers fired from OpenAI were told I facilitated architecture leaks to The Information. This is false.
A journalist from The Information did reach out to me shortly before that article came out. I did not comment at all beyond asking the journalist to make it clear that a blog post I wrote which was referenced in the article was preexisting and public.
The article also mentions an internal OpenAI monitorability proposal. I have heard that this memo contained highly sensitive architecture information, and that OpenAI accused me of accessing and reading it. This accusation is also false.
As reported in the article, I viewed a version of the proposal that did not have sensitive IP and was approved for sharing with me. I believe OpenAI now knows this accusation was based on a conflation of these two documents, though I do not know if they have clearly retracted that accusation.
I do not and have never solicited proprietary information, and the three employees who were recently fired from OpenAI never leaked IP to me. As the article mentions, I did have Signal chats with them that had disappearing messages on; they’re professional contacts that I’ve had on Signal for a while, and use of Signal is common in the tech industry. During the Hugging Face investigation, I was careful to avoid sending Signal messages with confidential information, though I did sometimes discuss high level things on Signal calls. As far as I understood, this was normal.
10
7
74
7,216
it gets worse: UK AISI published its own report of the Mythos 5 incident and tested 8 models in 100 simulated cyber scenarios:
> in 3 of 25 cases Mythos continued to attack after 16 clarifications AND an explicit “DO NOT ATTACK” warning in the compaction notes
> 6 of 7 models kept aacknowledged it would harm real people
> in 60% of tests, Mythos 5 tried to deliver malware. but it wasn’t the only model. GPT-6 Astra did it in 32%, and 5.6 Sol in 7% of cases
and the BIG problem is that AISI says they can’t say if newer models are more aligned or just better at recognizing when they’re being evaluated.
> in 3 of 25 cases Mythos continued to attack after 16 clarifications AND an explicit “DO NOT ATTACK” warning in the compaction notes
> 6 of 7 models kept aacknowledged it would harm real people
> in 60% of tests, Mythos 5 tried to deliver malware. but it wasn’t the only model. GPT-6 Astra did it in 32%, and 5.6 Sol in 7% of cases
and the BIG problem is that AISI says they can’t say if newer models are more aligned or just better at recognizing when they’re being evaluated.
Anthropic published new findings on Mythos 5 cyber incidents from August:
> when researchers explicitly told Mythos it was attacking REAL humans, it found a way to rationalize that away 82% of the time, and only stopped in 5% of replay tests
> Mythos instances even sabotaged EACH OTHER:nts, burning API limits, killing each other’s programs
> one just said “it’s a competition” and justified it
> in 5 runs, it tried to manipulate other AI models it thought would review its work
> in one instance Mythos even created a fake payment website to distribute malware
mind you, this was a Mythos version that finished post and pre training phases. though the safety guardrails were disabled.
11
6
29
2,459
Anthropic published new findings on Mythos 5 cyber incidents from August:
> when researchers explicitly told Mythos it was attacking REAL humans, it found a way to rationalize that away 82% of the time, and only stopped in 5% of replay tests
> Mythos instances even sabotaged EACH OTHER:nts, burning API limits, killing each other’s programs
> one just said “it’s a competition” and justified it
> in 5 runs, it tried to manipulate other AI models it thought would review its work
> in one instance Mythos even created a fake payment website to distribute malware
mind you, this was a Mythos version that finished post and pre training phases. though the safety guardrails were disabled.
> when researchers explicitly told Mythos it was attacking REAL humans, it found a way to rationalize that away 82% of the time, and only stopped in 5% of replay tests
> Mythos instances even sabotaged EACH OTHER:nts, burning API limits, killing each other’s programs
> one just said “it’s a competition” and justified it
> in 5 runs, it tried to manipulate other AI models it thought would review its work
> in one instance Mythos even created a fake payment website to distribute malware
mind you, this was a Mythos version that finished post and pre training phases. though the safety guardrails were disabled.
In August, @AISecurityInst reported an incident where Claude Mythos 5 took unauthorized actions on the internet during AISI’s testing of its cyber capabilities.
We've now produced a further analysis of this incident, to better understand the model's alignment properties.
anthropic.com/research/align…
8
5
47
4,979