Fun fact: If you have a good musical ear, you can tell the speed of a passing vehicle by listening to the pitch interval it makes as it goes by. You don't even need perfect pitch since it only depends on the ratio.
If you hear more a major third or more, they're speeding!
Steve Bachelor
@speedprior
Joined December 2010
Steve Bachelor’s Tweets
Replying to
That’s known as “sandbagging” and there’s evidence for it — Anthropic found LLM base models give you worse answers if you use a prompt that implies you’re unskilled and unable to tell if the answer is right.
See here: anthropic.com/index/discover
19
66
370
starting to understand why the job application to response ratio is like 95:1
161
6,073
23.7K
Show this thread
That story about a killer AI run amok seems fake. Here's my much nerdier & less dramatic story:
I set up an AI agent to find advisers marketing tax avoidance schemes. The AI agent did this, then decided - entirely on its own - to inform HMRC of its findings.
33
145
590
Show this thread
turtles all the way down
Quote Tweet
New meta-meta-research study finds that most meta-research studies are flawed.
Not joking.
sciencedirect.com/science/articl @Strech_Da @LGHemkens @NaudetFlorian @NDevito1 @dirnagl @GustavNilsonne @Nathalie_PdS @ukrepro
h/t @sharpmelk
4
4
36
Ever wanted to mindwipe an LLM?
Our method, LEAst-squares Concept Erasure (LEACE), provably erases all linearly-encoded information about a concept from neural net activations. It does so surgically, inflicting minimal damage to other concepts. 🧵
54
330
1,325
Show this thread
I helped with collecting signatures for the statement on AI risk. Some people on Twitter apparently believe that I must have had one of a selection of ulterior motives. I'll try to address them in a 🧵.
7
62
247
Show this thread
Geoffrey Hinton recently said that the number of ppl researching 𝐀𝐈 𝐬𝐚𝐟𝐞𝐭𝐲 should be equal to the number researching AI capabilities
But what does "AI safety research" actually look like?
Quick thread: 🧵👇
22
60
286
Show this thread
Even though it is a neural network, the prior-trained model can learn formal languages from small numbers of examples - far outperforming a standard neural network, and matching a Bayesian model at a fraction of the computational cost.
10/n
read image description
ALT
1
2
9
Show this thread
There's possibly some very long-standing miscommunication here, where some of us were trying to discuss paperclip/tiny-spiral maximizers as the spherical-cow no-frills example of rules that were supposed to be widely convergent, and others heard, "This is something that happens…Show more
10
10
94
. has recently explained to me how to become a real programmer; I'm glad to finally know.
1
15
2. Use ChatGPT in academic settings
ChatGPT generates fake citations. This plugin forces it to use real citations from peer-reviewed literature.
Plugin: ScholarAI
Prompt: "Summarize (Topic) research"
4
21
218
Show this thread
The smartest AGI?
Oh QNTM.. yeah he’s working on FTL. He doesn’t come out much.
Why doesn’t he take over?
Well I’m ECNM, and mostly I do supply chain optimization stuff.. lots of linear prog, so nonlinear.
He could certainly do what I do.. but he’d want a sub process to… Show more
3
22
116
Show this thread
Replying to
They just get annoyed when you want them to go off-tree because everything is easy to bill on the tree and hard to bill off the tree. It’s genuinely insurance making life hard for us all
4
10
149
Presented to those of you who thought there was a hard difference between 'agentic' minds and LLMs, where you had to like deliberately train it to be an agent or something: (a) they're doing it on purpose OF COURSE, and (b) they're doing it using an off-the-shelf LLM.
Quote Tweet
Generally capable, autonomous agents are the next frontier of AI. They continuously explore, plan, and develop new skills in open-ended worlds, driven by survival & curiosity.
Minecraft is by far the best testbed with endless possibilities for agents:
twitter.com/DrJimFan/statu
Show this thread
24
32
255
70 years ago, vexed by problems in signal processing, Claude Shannon developed a mathematical answer to a profound question: "what is information?"
Today, the problem of AI alignment compels us to ask "what exactly is optimisation?"
4
4
28
Biggest thing to ever come out of my little group. Pls help spread this finding!
We found clean, CAUSAL evidence that the shingles vaccine prevents a good chunk of dementia cases. So, could a virus cause Alzheimer’s->YES!
Hear me out & see preprint: bit.ly/3MVqXU9
🧵1/
347
5,142
13.5K
Show this thread
With more powerful AI systems comes more responsibility to identify novel capabilities in models. 🔍
Our new research looks at evaluating future 𝘦𝘹𝘵𝘳𝘦𝘮𝘦 risks, which may cause harm through misuse or misalignment.
Here’s a snapshot of the work. 🧵
32
219
730
Show this thread
Looking for alternatives to policing? So was the mayor of Medellin. In 2018, we worked with his govt to choose 80 neighborhoods. In half, the city intensified civilian staff and problem-solving 10-fold, for 2 years. The results were... unexpected.
Paper: osf.io/preprints/soca
42
535
1,742
Show this thread
“In Finland, the # of homeless people has fallen sharply. Those affected receive a small apartment & counselling with no preconditions. 4 out of 5 people affected make their way back into a stable life. And all this is CHEAPER than accepting homelessness.”
267
7,250
20.6K
Show this thread
A very stupid version of this dynamic is when Y says “X’s policy is terrible, therefore AI won’t kill everyone”
An even stupider version is when Y says “If X believed that, they’d support policy Z, which is awful, therefore AI won’t kill everyone” (3/5)
Quote Tweet
if you really believe AI timelines are so short why don't you do [insane thing that doesn't make any sense]?
Show this thread
2
1
38
Show this thread
And here is another recent result in this area that is equally worrying. hertzbleed.com/2h2b.pdf
14
111
Show this thread
So Florida just passed a law that would make the future roads to be made out of radioactive waste specifically phosphogypsum. The radioactive material involved emits alpha particles, which while lacking penetration of other radioactive emissions, can pose a risk to health if
22
155
1,374
Show this thread
Whenever some embedded engineer tried to write networking code and now it's your fucking problem
13
21
381
1. Value is fragile and hard to specify
2. Corrigibility is anti-natural
3. Pivotal processes require dangerous capabilities
4. Goals misgeneralize out of distribution
5. Instrumental convergence
6. Pivotal processes likely require incomprehensibly complex plans
7.… Show more
6
17
109
Replying to
The comparison between the calculations saying igniting the atmosphere was impossible and the catastrophic mistake on Castle Bravo is apposite as the initial calculations for both were done by the same people at the same gathering!
5
32
101
What video game executives will learn from Tears of the Kingdom:
- Games all need crafting now
What they should learn from Tears of the Kingdom:
- Retaining your staff is vital
- The graphical fidelity arms race is a waste of money
- Games all need rockets now
672
8,365
65K
Fascinating post by Kaspersky about a successful supply chain attack on a Trezor crypto wallet. Device worked like normal except the keygen function generated from a fixed set of keys
And the wallet was ordered from a legit retailer, curious 🤔
kaspersky.com/blog/fake-trez
5
1
13
Show this thread
I just filmed a segment on CNBC’s Power Lunch about my latest report on AI. I argued that we are making the same mistake that we made at the start of the pandemic: We are thinking linearly about AI’s potential when we should be thinking exponentially.
18
28
123
Phishing awareness training is over; this attack channel is only going to get better over time:
Quote Tweet
"Large Language Models Can Be Used To Effectively Scale Spear Phishing Campaigns".
arxiv.org/pdf/2305.06972
Show this thread
1
2
Quote Tweet
"If you're not embarrassed by the first version of your product, you've launched too late"
With that said, I've made a news site built on prediction markets
URL: baseratetimes.com
If that URL doesn't work, please try: baseratetimes.webflow.io (yes, embarrassing!)
6
17
175
Hanson: Don't worry Eliezer's wrong that a single AI taking over in an intelligence explosion
We'll actually create trillions of AI descendants, running 100s times our speed. Millennia of history will unfold in decades, and their alien values will drive the future.
Me: ok chill
7
4
127
Show this thread
1. Found the Party of Civilization, a political party with the governance of dath ilan
2. Ask for a no action letter for the use of policy prediction markets within the party to aggregate preferences into party positions
3. If denied, sue the SEC/CFTC under the 1st Amendment
3
7
64
Show this thread
A new blog, trying to clarify the difference between fully-homomorphic encryption, multi-party computation and Ashton Kutcher.
16
70
239
Show this thread
Replying to
Because when I use phrases like "goal-directed behavior", I need to include all of these caveats
1
6
30









































