We should never forget just how much power GPUs to power our AI can cost, so use it wisely. Of course making a search on Google also costs power, but definitely an order of magnitude less. :)

Trust, but verify... :)

Let's hope AI brings us to fusion a bit quicker :)

Maybe sometimes someone should tell us no. :)

Increased ease of automation in economically risky times leads to dangerous temptation, moreso if laying off people then causes a drop of consumption while overestimating AI may lead to productivity problems.

I guess it's not like they don't already have access to everything... :) We are big fans of the cloud providers, but sometime it is worth thinking about how hostile the internet has become to having full control over your own data. Just try setting up your own mail server and passing a spam check.

Even though most people don't come into touch with coding, we should not underestimate just how much the impact of AI has now on software development and research. Impact which can both be good and bad depending on how careless how the user is. :)

Many of the better models have wide knowledge and much faster response than a human. They make mistakes, but who doesn't. In many cases it's good memory management that helps them keep a consistent personality and path to solving a problem which is something you can also improve software-side.

The question is whether disclaimers and basic guardrails stop the user from getting this information or just makes it less predictable.

The economy of scale gives the big companies an edge, still they should be careful assigning fixed prices to varying usage of products in an evolving market. :)

Until the beast can escape into our phones or through robotics the immense energy and hardware needs of giant AI should make it difficult to get out of control. And while we do now have agents, ones based on our best models are still rather slow. For now let's worry about humans abusing AI. :)

Very important initiative, but we do worry a bit about it's feasibility. Watermarks can often be filtered out by running an open-source model afterwards, meaning you can likely control and punish legal entities, but it may be different to fight malicious actors.

It's beautiful :) though in the end it shows that the molecules just differ by limited functional groups, which may limit the search space we can cover. AI can help us speed up simulations that help us explore entirely new molecular families which we may never find ourselves. :)

Realistic LLM security will be essential for all applications. Even subtle influences on peoples mental state may in fact be much more dangerous than telling people how to make nuclear weapons...