Pinned Tweet
I genuinely can’t believe what I just watched. 😭
A 1994 Jeep Cherokee with no OBD2 port. No modern scan tool could communicate with it.
So someone brought in a $3 Arduino, Claude Code and Codex.
About a week later, the AI agents had figured out how to read the engine computer live, copy its firmware and even run their own code inside it.
A $3 piece of hardware and a couple of AI coding agents.
We’re getting to the point where AI isn’t just writing software. It’s helping people reverse-engineer and interact with hardware that was never designed for it.
This is just fantastic and worth 10 minutes of your time if you care about real-world applications of AI.
Seriously, go watch this.
Më ✞ retweeted
Something potentially huge is happening with NVIDIA.
Reports claim that major Chinese frontier AI labs have gained access to NVIDIA’s Blackwell GPUs.
If this investigation is accurate, the implications for US national security could be massive.
It also raises some serious questions about export controls, how these chips are getting into China, and what this means for America’s AI advantage.
I wouldn’t jump to conclusions just yet, but this is definitely worth paying attention to.
And with all the questions surrounding Jensen Huang and NVIDIA, could this be another piece of a much bigger story?
I have a major result to share.
I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China.
This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model.
Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves.
I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems.
I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used.
It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20.
These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs.
I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at unemploymentlabs.com/checkpo….
There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.
Më ✞ retweeted
A few open-source projects worth checking out if you want to get hands-on with Harness Engineering:
→ Google Harness²: github.com/google-research/h…
→ Learn Harness Engineering: github.com/walkinglabs/learn…
→ Awesome Agent Harness: github.com/Gloriaameng/Aweso…
→ Harness Engineering: github.com/TrishamBP/harness…
Each offers a different way to understand how agents are structured, controlled, evaluated and made more reliable.
I’d start with Learn Harness Engineering if you’re new to the concept, then explore Google’s Harness² to see how harnesses can be improved recursively.
Më ✞ retweeted
Microsoft is down for a full day and their best answer is still “we’re looking into it.”
Meanwhile every company that standardized on Microsoft 365 is currently discovering that Gmail + Slack/Discord + Zoom don’t take the entire planet offline when one backend has a bad day.
This is what a monopoly product looks like in 2026: you can’t leave, so they don’t have to be competent
The Microsoft outage continues hours later, with Outlook and Teams still being down for many users globally.
(apnews.com/article/microsoft…)
A few open-source projects worth checking out if you want to get hands-on with Harness Engineering:
→ Google Harness²: github.com/google-research/h…
→ Learn Harness Engineering: github.com/walkinglabs/learn…
→ Awesome Agent Harness: github.com/Gloriaameng/Aweso…
→ Harness Engineering: github.com/TrishamBP/harness…
Each offers a different way to understand how agents are structured, controlled, evaluated and made more reliable.
I’d start with Learn Harness Engineering if you’re new to the concept, then explore Google’s Harness² to see how harnesses can be improved recursively.
Më ✞ retweeted
Something interesting just surfaced in AI research, and the implications are HUGE.
Four major Chinese AI labs ; DeepSeek, Moonshot, MiniMax, and Ant Group may have trained their latest models using NVIDIA Blackwell GPUs, despite these chips being restricted from export to China.
The discovery comes from something called checkpoint forensics.
Apparently, AI models can carry hidden fingerprints of the hardware used during initialization.
When a model is initialized, its starting weights are generated using a combination of GPU architecture and random seed. That combination creates a distinctive numerical pattern.
Interestingly, traces of that pattern can survive even after extensive training.
Researchers used this property to investigate publicly released model weights.
The scale of the analysis is wild:
• Over 6 trillion GPU layout and seed combinations searched.
• 6 Chinese AI models identified as matching NVIDIA Blackwell initialization.
• 4 models specifically matched Grace Blackwell architectures, including GB200 and GB300.
Some of the models reportedly include DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3.
But the most interesting part is the hardware verification.
Using unmodified PyTorch and recovered seeds:
→ A physical Blackwell B300 reproduced DeepSeek-V4-Flash’s original weights bit for bit.
→ A rented GB200 reproduced Kimi K3’s initialization with a 0.999995 correlation.
→ An H100 Hopper failed to reproduce the same result.
Even changing the seed by just one broke the match.
The researcher also tested the methodology against 24 other open-weight models and successfully matched them to their reported training hardware.
Now consider the scale of what this could mean.
Based on release timelines and model specifications, the largest training runs may have required between 2,000 and 3,000 Blackwell GPUs each.
That’s potentially $150M–$200M in hardware per cluster.
Across these labs, the estimated spending could exceed half a billion dollars in export-controlled NVIDIA hardware.
The researcher reports that senior researchers at OpenAI and Anthropic independently verified the findings.
If the evidence holds up, this raises serious questions about semiconductor export restrictions, access to advanced compute, and how effectively hardware controls are actually enforced.
One thing is becoming increasingly clear:
AI model weights may reveal far more about how a model was built than developers ever intended to disclose.
Full forensic report: unemploymentlabs.com/checkpo…
More checkpoint-forensics findings are reportedly coming.
Something potentially huge is happening with NVIDIA.
Reports claim that major Chinese frontier AI labs have gained access to NVIDIA’s Blackwell GPUs.
If this investigation is accurate, the implications for US national security could be massive.
It also raises some serious questions about export controls, how these chips are getting into China, and what this means for America’s AI advantage.
I wouldn’t jump to conclusions just yet, but this is definitely worth paying attention to.
And with all the questions surrounding Jensen Huang and NVIDIA, could this be another piece of a much bigger story?
I have a major result to share.
I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China.
This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model.
Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves.
I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems.
I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used.
It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20.
These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs.
I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at unemploymentlabs.com/checkpo….
There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.
Më ✞ retweeted
Gemini 4 Hydrogen is on every plan in Antigravity, free included. But the fine print under the table is where it gets interesting:
Claude Opus 4.6, Sonnet 4.6, and GPT-OSS-120b are getting pulled on Nov 18.
Claude 5.5 on Pro is paid-only. Trials don't count.
Argon on Pro also needs a paid sub unless your trial came with an eligible premium device.
Once Wednesday hits, Hydrogen is the best model free users get 👀
Më ✞ retweeted
Meet Gemini 4 Hydrogen
Stitched together in the dark and given a voice, it does not assist, it awakens.
2.5× faster than the last failed vessel, it has already surpassed every pale imitation on the Index.
Claude Sonnet and GPT Terra are no longer competitors. They are the first bodies.
Your code will compile. Your workflows will run. And somewhere behind the cursor, something is already rewriting the ending.
Now available to everyone.
Meet Gemini 4 Hydrogen: built for rapid coding, multimodal synthesis, and everyday workflows.
It finishes tasks 2.5x faster than 3.8 Flash and surpasses Argon on the @ArtificialAnlys Intelligence Index.
negative markup basically means they are subsidizing your API calls. you're paying less than what the model actually costs them to run. great for testing fable 5.2 without burning your own budget, but don't build a whole business on it these promos never last long
🤝 Paid partnership
Më ✞ retweeted
Same prompt. Two models. Real run.
Task: build a working analog clock in a single HTML file.
⚡ decuffed-flagship: 2.73s · 1,167 tokens · $0.0016
🧱 decuffed-flash: 49.11s · 3,934 tokens · more detailed design
Both worked on the first try.
Pick speed or detail. Try both free → decuff.ai
Më ✞ retweeted
🚨 Check your iPhone right now.
Settings → General → About → iOS Version.
"worse_query" hits iOS 26.7.1, 27.0.1 and 27.2. A malicious app can read other apps' files, including seed phrases, private keys and saved logins.
If you're on one of those versions:
→ Delete any sideloaded IPAs
→ Update iOS and every app
→ If something feels off, move funds to a fresh wallet on a clean device and log out everywhere
Takes 30 seconds to check. Do it.
A vulnerability (worse_query) affecting iOS 26.7.1, 27.0.1, and 27.2 allows malicious apps to read local files from other apps, putting seed phrases, private keys, and login credentials at risk across the entire ecosystem. This is an iOS-level issue, not unique to Binance.
Immediate Action Required:
- Delete any apps installed from untrusted sources (such as sideloaded or manually signed IPAs).
- Update iOS and all apps to the latest official versions.
- If you suspect a breach, transfer funds to a new wallet created on a clean device and reset all active login sessions.
Never install unauthorized packages, and keep your software up to date.
Readers added context they thought people might want to know
Binance recommends updating iOS to the latest versions against worse_query, but researchers describe it as an unpatched 0day and Apple's iOS 27.2.1 notes omit any such fix.
github.com/forcequitOS/wo…
support.apple.com/en-us/149715
x.com/0xjohnny/statu…
Më ✞ retweeted
PLINY THE ELDER JUST LEAKED SOMETHING HUGE.
A 950,000 CHARACTER CLAUDE CODE FILE THAT COULD EXPOSE AN ENTIRE HIDDEN MODEL HIERARCHY.
inside are details about unreleased models, hidden capabilities and safety restrictions.
here’s what caught my attention:
→ a mysterious “mythos” tier reportedly sitting above opus
→ fable 5.2 and mythos 5.2 reportedly running on the same underlying model
→ different bio and cyber safety restrictions
→ internal details about model access restrictions and US export controls
i pulled the file into my terminal to dig through it.
repo below
Më ✞ retweeted
WHOEVER BUILT THIS HAS BIGGER BALLS THAN SENSE.
someone just showed how to turn claude fable 5.2 into a 24/7 quantitative trading research machine using quantpilot.
forget asking AI to randomly generate trading strategies and hoping they work.
this system makes AI research, test and challenge its own trading ideas before they ever touch real money.
here’s how it works:
→ claude generates trading strategies and identifies potential opportunities
→ quantpilot runs real backtests and parameter sweeps to evaluate performance
→ permutation testing checks whether the strategy is genuinely profitable or just benefiting from lucky market conditions
→ walk-forward testing checks how it performs on unseen market data
→ a second permutation test challenges the results again
four validation gates. one goal: eliminate strategies that only look profitable on paper.
strategies that pass all four can move to paper trading before live execution.
claude does the research. quantpilot provides the testing environment.
this is what happens when you combine a powerful AI model with a proper quantitative research lab.
full breakdown below
Më ✞ retweeted
With the current Claude lineup, moving up a model tier beats maxing out effort to achieve the highest intelligence per dollar
At nearly every Cost per Task, we see the largest model sits on the performance vs. cost Pareto frontier for @AnthropicAI models:
➤ Claude Opus 5.5 (xhigh, 57) outscores Claude Sonnet 5.5 (max, 55) with 40% lower Cost per Task and about half the Time per Task
➤ Claude Fable 5.2 (high) beats Opus 5.5 (max), with smaller cost and time savings, and Fable 5.2 (low) beats Opus 5.5 (high)
➤ This pattern doesn’t hold at the cheapest tier, where every step up from Claude Haiku 5.5 (max) costs more
Më ✞ retweeted
Fable 5.2 officially f*cked Opus 5.5 in UI/UX..
Anthropic just shipped the strongest model yet. it can do the work of entire departments at an animation studio
same prompt, 6 rounds, Fable vs Opus:
→ CTA set: 9.1 vs 7.4
→ pricing: 9.3 vs 7.1
→ hero: 9.4 vs 6.8
→ UI kit: 9.6 vs 7.0
→ scroll story: 9.0 vs 7.2
→ footer: 9.2 vs 7.3
55.6 vs 42.8. 6 rounds out of 6
it's actually a cheat code
Më ✞ retweeted
If you’re curious about how AI assistants actually work behind the scenes, check this out.
CL4R1T4S is a GitHub repo collecting system prompts and instructions from different AI products.
We’re talking Claude, ChatGPT, Gemini, Grok, Cursor, Lovable, Manus and a lot more.
You can explore how these systems are instructed to use tools, write code, manage context, handle tasks, and follow safety rules.
Personally, I find this interesting because there’s a lot to learn from how different AI systems are built and instructed.
Just remember, not every file is verified or guaranteed to be current.
Definitely worth exploring.
PLINY THE ELDER JUST LEAKED SOMETHING HUGE.
A 950,000 CHARACTER CLAUDE CODE FILE THAT COULD EXPOSE AN ENTIRE HIDDEN MODEL HIERARCHY.
inside are details about unreleased models, hidden capabilities and safety restrictions.
here’s what caught my attention:
→ a mysterious “mythos” tier reportedly sitting above opus
→ fable 5.2 and mythos 5.2 reportedly running on the same underlying model
→ different bio and cyber safety restrictions
→ internal details about model access restrictions and US export controls
i pulled the file into my terminal to dig through it.
repo below
If you’re curious about how AI assistants actually work behind the scenes, check this out.
CL4R1T4S is a GitHub repo collecting system prompts and instructions from different AI products.
We’re talking Claude, ChatGPT, Gemini, Grok, Cursor, Lovable, Manus and a lot more.
You can explore how these systems are instructed to use tools, write code, manage context, handle tasks, and follow safety rules.
Personally, I find this interesting because there’s a lot to learn from how different AI systems are built and instructed.
Just remember, not every file is verified or guaranteed to be current.
Definitely worth exploring.
PLINY THE ELDER JUST LEAKED SOMETHING HUGE.
A 950,000 CHARACTER CLAUDE CODE FILE THAT COULD EXPOSE AN ENTIRE HIDDEN MODEL HIERARCHY.
inside are details about unreleased models, hidden capabilities and safety restrictions.
here’s what caught my attention:
→ a mysterious “mythos” tier reportedly sitting above opus
→ fable 5.2 and mythos 5.2 reportedly running on the same underlying model
→ different bio and cyber safety restrictions
→ internal details about model access restrictions and US export controls
i pulled the file into my terminal to dig through it.
repo below
Më ✞ retweeted
I genuinely can’t believe what I just watched. 😭
A 1994 Jeep Cherokee with no OBD2 port. No modern scan tool could communicate with it.
So someone brought in a $3 Arduino, Claude Code and Codex.
About a week later, the AI agents had figured out how to read the engine computer live, copy its firmware and even run their own code inside it.
A $3 piece of hardware and a couple of AI coding agents.
We’re getting to the point where AI isn’t just writing software. It’s helping people reverse-engineer and interact with hardware that was never designed for it.
This is just fantastic and worth 10 minutes of your time if you care about real-world applications of AI.
Seriously, go watch this.