Pinned Tweet
Replying to @skwp
Here's the thing you need to think about. If AI is so smart and so useful, why are we even asking it to write software to solve problems for us? Shouldn't the AI just do the thing we need?
1
3
300
Bryan Murdock retweeted
2 minutes left in this BYU vs Kansas game, and there have been... zero penalties?
Has a zero-penalty game EVER happened?
Anthropic's IPO bank closed position meant to stabilize stock, removing source of support reut.rs/4zD6AjM
Bryan Murdock retweeted
We have finally built The One Ring from seminal literary classic The One Ring Must Be Destroyed.
I feel like I should be getting paid a consulting rate to teach AI how to do my job for my employer
1
2
30
Here's my conspiracy theory:
1. Create chatbot
2. Raise lots of money based on chatbot potential to be AGI
3. When chatbot progress slows, hype up how, oops, hackers can use it!
4. A bunch of outages occur because chatbot not that good
5. Blame hackers
1
1
42
Bryan Murdock retweeted
"We have met the enemy and he is us."
This cracked me up, @SPACEdotcom :) Thanks for protecting me from … me.
Bryan Murdock retweeted
Next time someone tell you that subscription AIs don't hallucinate at all anymore or have greatly reduced the problem - ask them to Google it. Or give them these stats with citations :
According to the First Page Sage 2026 Report, on subscription models Claude has a response-level hallucination rate of 8.3% and ChatGPT sits at 11.0%,
Volatile real-world details. Questions about local business hours and addresses trigger hallucinations 26.0% of the time, while breaking current events sit at 21.0% and specific product pricing sits at 19.4%.
Across major platforms, queries about niche subjects produce a 31.0%hallucination rate, making them 4.4 times more error-prone than simple factual lookups at 7.0%. Questions containing a false premise are the worst performers, triggering fabrications 46.1% of the time as models play along with the user's mistake.
Obscure public figures and trivia. OpenAI's own system cards show that answering short factual trivia from memory causes hallucinations 40% to 51% of the time on flagship reasoning models, while asking about lesser-known public figures triggers errors 33% of the time.
High-stakes domains. General legal questions produce fabricated statutes or case law 23.5% of the time, though specialized audits of federal court cases have recorded rates as high as 58%. Financial queries average a 16.8% error rate, medical and health queries sit at 15.0%, and software coding queries average 12.1%.
When tested on 6,000 difficult factual questions without internet access on the Artificial Analysis AA-Omniscience Leaderboard, hallucination rates rise dramatically because models guess when they miss. OpenAI's GPT-6.1 Terra hallucinates on 34.5% of its non-correct answers, GPT-6 Astra hits 49.8%, and GPT-6.1 Sol reaches 51.9%. Anthropic's Claude Sonnet 5.5 hallucinates on 43.3% of misses, Claude Opus 5.5 hits 50.3%, and Claude Fable 5.1 reaches 70.0%despite achieving the highest raw accuracy.
suprmind.ai/hub/ai-hallucina…
Bryan Murdock retweeted
This has been a pretty eye-opening week for me. Not so much because of the revelations about certain people in tech and business but the reactions to them.
It's surprising how quickly people want to overlook some pretty serious ethical questions because of what someone has achieved.
I'm a naturally optimistic person, but I increasingly feel the US has lost its way. And I don't think this is a left or right issue. We've become so obsessed with individual success that moral responsibility seems to have become optional.. and we're trying hard (with some success) at exporting this thinking abroad.
Individualism is one of America's greatest strengths, but somewhere along the way we've confused personal freedom with freedom from responsibility to others. I don't think a society can flourish for long when success becomes an excuse for anything.
Maybe there's something more fundamentally positive at play here than I understand, but I recently read Simon Blackburn's Think and this part felt particularly relevant:
"A society in which people were incapable of recognizing the others as having a complaint, whatever they do, would be one without an ethic -- but for that very reason it would be hard to recognize it as a society at all."
Bryan Murdock retweeted
The FAA has lifted the nationwide ground stop, which was originally imposed at 3:54 p.m. Eastern Time, ending several hours of confusion and operational disruption across the aviation industry.
However, lingering delays and disruptions are likely to persist as airlines and airports work to restore normal operations. A full recovery may take time.
Bryan Murdock retweeted
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.
The thing I hate most about the AI craze is the excitement that we don't (well, won't, because we ain't even there yet) have to learn anything anymore. AI will understand and figure everything out for us. Complete removal of my agency is not a selling point for me.
Bryan Murdock retweeted
If AI CEOs really thought the world was ending, would they be taking their companies public and building data centers? Steve Eisman says the doomsday talk is about getting governments to regulate the competition away. He gives it a 0% chance.
Watch the full video for the scenarios he puts odds on, including this one at zero. Link in bio.
How is this approach not begging these opaque proprietary LLMs to insert bugs and backdoors into your product? Have we all forgotten every single lesson we've learned about software security?
I used to be a language lawyer. In the late 80s I _studied_ the ARM (Annotated Reference Manual for C++). I knew the language cold, inside and out. I could make it do backflips. I shipped a lot of code using it. A LOT.
I’ve attacked other languages with a similar intensity and depth. Until recently, that was simply the only way to become adept at building software systems.
I now consider the deep understanding of a language to be a supreme waste of time and effort. I don’t want to learn Rust, or Golang, or Typescript, or any of the thousands of other languages out there. I’ll leave that to my agents.
What I need now is to understand the high level gestalt of the system. I want to see the systems and subsystems. I don’t need to inspect the cogs. If I truly need to understand a detail at the cog level, I’ll interrogate my agents. I’m not going to pretend to be a language expert anymore.
1
34
























