Took me a while to go through this paper introducing "Epiplexity", but this is really one of the best ones I've read in a while.
It introduces a simple, new theoretical angle which explains a lot of the information theory paradoxes in ML - and can have practical impacts too.
arxiv.org/abs/2601.03220
Learn more in our blog post, "The Scientist AI: Safe by Design, by Not Desiring":
lawzero.org/en/unlisted/...
(4/4)
At LawZero, we're rethinking the building blocks of frontier AI to create an intelligent machine that is both highly capable and safe-by-design. We’re excited to share our first blog post outlining some of the objectives and core components of our Scientist AI project. 🧵
(1/4)
SafeRL techniques are usually required when you need to strengthen this learning mechanism because 1) you have to be really sure constraints will be respected 2) the learning mechanism from very strong penalties is not stable or 3) you want to learn from as few examples of failure as possible
If you have a couple of trajectories in a replay buffer where the agent's existence ends, usually these don't get rewards anymore. It is a state with a low value. So you would learn to avoid this - which is the same thing as fear.
If the algorithm instance has an objective and is smart enough to understand that its non-existence leads to failure on that objective, then the algorithm will fear for its existence.
LawZero is accepting applications as part of the SPAR Spring 2026 program!
If you're interested in studying model awareness or emergent misalignment, you can learn more and apply here: sparai.org/projects/sp26/.
Applications are open until Jan 14, 2026.
Hi RL Enthusiasts!
RLC is coming to Montreal, Quebec, in the summer: Aug 16–19, 2026!
Call for Papers is up now:
Abstract: Mar 1 (AOE)
Submission: Mar 5 (AOE)
Excited to see what you’ve been up to - Submit your best work!
rl-conference.cc/callforpaper...
Please share widely!
LawZero is accepting applications as part of the MATS Summer 2026 program!
If you're interested in contributing to LawZero's Scientist AI program and studying model awareness, you can learn more and apply here: www.matsprogram.org/stream/lawzero
Applications are open until Jan 18, 2026.
To be honest, I hope you are right. I believe that there is high economic value in current and slightly better AI once correctly integrated. But there are many risks with exponential AI improvements, and I hope your argument of lack of economic sustainability stands and this progress slows down.
I also believe AI's current costs can and will be reduced significantly (per ill-defined unit of performance), the current models being non-optimal algorithms running on non-optimal hardware.
Every year, smaller models catch up to last year's champions. There is a lot of space to reduce costs.
Absolutely!
We are talking about what happens in the future - there are multiple assumptions and bets. I would not bet my house on everything I said ^^
Yet it seems to be a very plausible scenario.
Also, they have plenty of monetization schemes in the sleeves. Social media style/search engine style - gather data from users, make companies pay for their products to come out in the results, etc.
This is like the early internet. It will grow, become necessary, and then get monetized.
I think you may be underestimating the usage of these tools given for very low cost to other parts of the economy. Once everyone has found a way to use AI for value in their specific domain, the prices will get higher without hindering adoption.
LawZero is growing fast, and we're always looking for dedicated people to join our team.
If you're interested in working on technical safeguards to create safe-by-design AI systems, check out the openings on our website and don't hesitate to reach out to our team!
job-boards.greenhouse.io/lawzero
Mila's annual supervision request process is now open to receive MSc and PhD applications for Fall 2026 admission! For more information, visit mila.quebec/en/prospecti...
Establishing where we collectively draw red lines is essential to prevent unacceptable AI risks.
See the statement signed by myself and over 200 prominent figures:
red-lines.ai
You are right. I am sorry. I will remove my post.
It is very difficult, and stressful, to see the US going there. Being outside and therefore unable to do anything about it makes me feel powerless. I wish more would be done by those who can.
It is unfair to judge though. I am not in this position.
To friends worried about the US some months ago, I said things were OK as long as people like Kimmel could make fun of the president night after night in a large audience TV show.
This is not OK anymore. This country is running at fast speed into dark territory. Gotta wake up, American people. Now.
The article is also available on arXiv for when the Elsevier free sharing is over: arxiv.org/abs/2411.14618
It's among my favorite projects because 1) we used DL to reduce the costs of clean energy production 2) we managed to deal with many real-world and real-time constraints, and 3) it's fun to use neural nets to control enormous machines!
We ran this on a real turbine and achieved a 42% reduction of strain amplitude after only 7 startups, compared to the standard one. It's significant! We now aim at making it generally applicable for different types of turbines (this was for a Francis, Kaplans have an additional dimension).
Once the model is trained with enough data (or when we have used most of the startup budget), we finalize by optimizing the strain only. The parameters are always tested on the turbine, allowing to confirm the predictions with ground truth measurements.
The virtual sensors are integrated in a black box optimization loop, which also includes a turbine dynamics simulator. For several startups, we use active learning, with an ensemble and optimism under uncertainty, to output the parameters to run next to get the most informative data for the model.
This is where machine learning comes into play. We use a neutral networks to create a "virtual sensor" predicting the strain amplitude for a given turbine strain. It is trained on the trajectories we currently have measured.
The challenge is that the sensors for strain (which leads to fatigue and cracks) don't last long on such turbines. They get flushed away! We have ~10 tries only to find the best set of parameters which will reduce the strain amplitude.
The strain is what leads to fatigue and cracks. As every turbine is different, it's hard to know in advance which startup parameters lead to the lowest strain. We have to figure it out during the commissioning campaign, when the turbine is placed in the dam.
There are different ways to start a turbine. Startup parameters control the vane opening/rotation speed trajectory until the desired speed is reached. They lead to different strains on the turbine blades.
As renewable energies are integrated in the power grid, hydroelectric turbines are expected to be stopped and started more frequently. But startups are a major source of fatigue on the turbine blades and reduce turbines' lifetime.
