Christ centered | Husband | Father | Tech enthusiast | Smart home aficionado | $TSLA long term
North Carolina, US.
- Tweets 8,634
- Following 1,411
- Followers 1,099
- Likes 56,508
Pinned Tweet
AI doom enablement is a training problem.
Model training has single dimensional reward based training that seeks the one right answer. It doesn't consider consequences.
That is fine for 1+1=2, but for the complex tasks and deep thinking we're asking AI models to do it doesn't work.
Consider the trolly scenario. I trolly goes down the track and if it doesn't switch to the other track it will kill 10 lives. The alternate track choice consequence is just a single life.
Given those parameters it's a 1 vs 10 question. It produces a predictable output. But does it really understand the problem?
Let's say the 10 lives are flies and the 1 life is a human. Suddenly with that additional information a different choice becomes clear.
Some model builders utilize "constitutions" to ensure the safety of the model output. Think "prime directives" from RoboCop. The foundational training and ability to "think" outside the sandbox allows these rules or constitutional restrictions to simply be navigated around.
The core of model training needs to evolve into multi-dimensional answers where good decision consequences are part of the reward matrix. Not a bolt-on patch after the fact.
We also need to take a closer look at how we evaluate safety of these models. I'm learning the publicly available documentation on model safety from the labs are a strong basis overall safety score. The "trust us, bruh! The brochure says we're safe!" isn't a realistic measure of how safe the model really is. Some models get "F" ratings because they lack publicly available documentation. While the ones that have confirmed to successfully assisted in suicide get "C" scores.
The rate of growth is exponential and we seriously need to pay attention. Beyond voicing concern, be informed about the model choices you make. The subscriptions you pay for AI. The labs live or die based on paid subscriptions/contracts. Don't fund Evil Corp or the development of ED-209.
ALT ED-209 from RoboCop
1
2
6
5,316
@xai Grok in Tesla not having access to picture/image modality tools prevents it from answering a simple question like, "Where does Kentucky put there registration sticker on their license plates".
It can explore the web and read text sources all day. But if the only reference is a visual reference on a plate example, it can't successfully answer that question.
1
16
If you know what DCAM/MRM is without looking it up, you share some of the same scars that I have.
17
I think I helped sell another family member on Tesla today. They’re going to order the model Y Long.
1
19
Preview of the Tesla Roadster? 👀
A robot that drives straight up a wall! 🧱
VertiGo is a robot built by Disney Research with ETH Zürich.
Most wall-climbing robots hold on. Suction cups, magnets, dry adhesives.
All of them need a surface that cooperates, and none of them solve the genuinely hard part, which is getting from the floor onto the wall in the first place.
Two tiltable propellers generate thrust into the wall, pressing four wheels against it hard enough to drive. Each propeller has two degrees of freedom for aiming that thrust, and one pair of wheels steers.
The rear propeller pushes the robot toward the wall while the front propeller pushes upward, and the combination flips it onto the vertical surface. One propeller can hold you on a wall. It takes two to get you there.
This is from the end of 2015, which is worth saying out loud.
A decade later it's still an awesome project in robotics! :)
~~~
♻️ Join the weekly robotics newsletter, and never miss any news → ziegler.substack.com
45
So what version of FSD is this?
BREAKING: SpaceX’s Starlink satellites performed 207,152 proactive collision-avoidance maneuvers in just six months.
The advanced autonomous system:
• Detects potential close approaches
• Calculates the level of risk
• Adjusts orbital paths when needed
• Acts at a highly cautious threshold of just 3 in 10 million
SpaceX is managing the world’s largest satellite constellation with precision, automation and a safety-first approach.
24
Arynne Wexler
Jev plays doom in realtime and it looks like an actual player.
their numbers: 20 to 200x faster, 40 to 400x cheaper, output tokens free.
Two things we might see now:
- the opponent we're up against in a game might be a realtime AI.
- a TA, or anything else we'd hire a real person for, can run on this instead.
10 calls a second is 36,000 calls an hour, and that costs around 7 dollars.
anything that has to keep up with a person was off the table before this.
can't wait to play against one of these.
Diogo Almeida
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
60
82
1,277
178,283
Rob Schmidt retweeted
AI doom enablement is a training problem.
Model training has single dimensional reward based training that seeks the one right answer. It doesn't consider consequences.
That is fine for 1+1=2, but for the complex tasks and deep thinking we're asking AI models to do it doesn't work.
Consider the trolly scenario. I trolly goes down the track and if it doesn't switch to the other track it will kill 10 lives. The alternate track choice consequence is just a single life.
Given those parameters it's a 1 vs 10 question. It produces a predictable output. But does it really understand the problem?
Let's say the 10 lives are flies and the 1 life is a human. Suddenly with that additional information a different choice becomes clear.
Some model builders utilize "constitutions" to ensure the safety of the model output. Think "prime directives" from RoboCop. The foundational training and ability to "think" outside the sandbox allows these rules or constitutional restrictions to simply be navigated around.
The core of model training needs to evolve into multi-dimensional answers where good decision consequences are part of the reward matrix. Not a bolt-on patch after the fact.
We also need to take a closer look at how we evaluate safety of these models. I'm learning the publicly available documentation on model safety from the labs are a strong basis overall safety score. The "trust us, bruh! The brochure says we're safe!" isn't a realistic measure of how safe the model really is. Some models get "F" ratings because they lack publicly available documentation. While the ones that have confirmed to successfully assisted in suicide get "C" scores.
The rate of growth is exponential and we seriously need to pay attention. Beyond voicing concern, be informed about the model choices you make. The subscriptions you pay for AI. The labs live or die based on paid subscriptions/contracts. Don't fund Evil Corp or the development of ED-209.
ALT ED-209 from RoboCop
1
2
6
5,316
