Following the data behind AI so you can debunk big tech. Computer Engineering @ Brown | The AI Safety Guy Building a universal guardrail for your agents👇
Joined October 2026
- Tweets140
- Following47
- Followers39
- Likes52
Twelve correct AI answers became wrong because the answer key changed.
A simulated shopper wanted the cheapest suitable earbuds already in their order.
Twelve publicly saved AI runs chose the correct item: $232.49. But a revised benchmark answer key expected a $226.49 item that wasn’t in the order.
When we replayed those saved runs against that key, all twelve failed. We corrected one reference value and replayed the same runs. All twelve passed.
The AI responses hadn’t changed. The grading reference had.
This is a concrete example of how an answer-key mistake can make correct AI behavior look like failure.
This is just a preview of what we are building at millenniumresearch.ai/. Tap in if you want to learn more about AI safety 👇
39
Everybody is talking about Claude Sheets taking over Microsoft, but I think the real losers are third-party GPT learning wrappers.
Does anyone remember TurboAI and their ABG Marketing Videos? Yeah, why use Turbo's Docs when you can use Claude Docs.
Personally, I like this, even if I am skeptical about how much it will impact the top dog of Microsoft. Since GPT wrappers are becoming increasingly difficult to provide any value.
2
2
60
BREAKING: Australia considers world-first criminal charges for AI companies that fail to prevent harm
Is this the right step towards AI Safety, or placing unreasonable expectations on Big Tech?
abc.net.au/news/2026-11-11/a…
17
I keep getting interviews asking crazy questions about buckets and 4-liters, crossing rivers with chickens and foxes and seeds, and 8 balls with 2 scales.
I have a theory: AI makes any skill virtually untestable because people can learn anything in an hour. So the only thing interviewers test is how you THINK through problems.
22
Builders and engineers who HEAVILY use Codex/Claude/Grok Bot, what's one thing that you feel uncomfortable sharing with AI that you still do?
27
IF YOU NOTICED OPUS 5.5 BEING WORSE, ITS NOT A MISTAKE.
Anthropic seems to have been running experiments with YOUR model.
>If you had the ID header "34e", tool calls increased from 7 to 17.
>Some ID experiments with medium effort used as many tokens as HIGH EFFORT models.
So far 4 experiment IDs have been identified between October 10th and November 8th:
>dba, 34e, 763, and eab
1
43
China and the us are on opposite sides of the globe, are ideological polar opposites, yet their tech markets are crashing at the exact same time.
The obvious common thread is AI overvaluation, but is this crash only going to impact the US and China, or is it a foreshadowing of a larger global recession for countries that haven't heavily invested in AI?
8
Apparently, a startup founded by former DOGE members claims it hacked Anthropic in 12 hours, except that's what Anthropic wanted. 🤔
Cathedral was working through Anthropic's bounty program to find technical flaws in Anthropic's system for them.
The craziest part is that Anthropic ISN'T going to pay them because apparently, Cathedral hid additional reports they hacked into. 🤯
30
Apparently you can see people's X device type.
s=46 iPhone
s=20 Browser
s=19 Android
s=420 Elon's burner
Try it out with this tweet
29
DON'T GIVE AI YOUR SOCIAL SECURITY NUMBER
DON'T GIVE AI YOUR SOCIAL SECURITY NUMBER
DON'T GIVE AI YOUR SOCIAL SECURITY NUMBER
DON'T GIVE AI YOUR SOCIAL SECURITY NUMBER
DON'T GIVE AI YOUR SOCIAL SECURITY NUMBER
DON'T GIVE AI YOUR SOCIAL SECURITY NUMBER
1
21
