Welcome to the first weekly dAIgest! Our one stop overview of everything AI that happened this week. In this short 5-minute read, we'll cover the most important events and share our thoughts.
This week in bullet points:
Let's dive into it!
Even though we need a calculator to get the final price in the supermarket, we definitely haven't missed this one. OpenAI claimed it solved Navier-Stokes, a 19th-century fluid dynamics problem designated in 2000 as one of the seven Millennium Problems. The Navier-Stokes equation is used to model fluid movement using Newton's second law. The problem with fluids however, is that they can contain violent swirls, called turbulence. The specific issue lies in finding a mathematical proof that describes turbulence with infinite velocity or pressure. Introducing an infinite variable, "breaks" the equation and for mathematicians, that is mighty interesting.
Setting aside this layman's explanation, the research team at OpenAI used an internal model running for 88 hours to demonstrate the infinite variable break, after hearing a rumor on September 1st that a solution was found. It turned out later, that this rumor stemmed from the work of Levent Alpoge and Tristan Buckmaster. Many, including Buckmaster questioned if OpenAI benefited from the private research, which used LLMs provided by OpenAI. These comments sparked controversy on AI accelerated research and its implications. OpenAI denied having access to private chats, but "cannot rule out that de-identified data derived from their usage of our products helped improve our models". OpenAI also confirmed it will not claim the $1 million prize money. Whilst we don't want to take part in further speculation, the situation does raises questions about who or what can take credit when AI helps in these things. Nevertheless, we expect similar situations in the future as the models get better and better.
After last week's release of Astra, over the weekend pro-tier users of ChatGPT got access. And in similar fashion of all AI launches, the servers were surged with AI enthusiasts wanting to get their hands on the latest toy. The demand was so high, that on September 10th, OpenAI paused sign-ups specifically for their $200/month ChatGPT Pro tier, citing that these (power) users put the most strain on the system. This again shows the major interest in Astra, and we are all for it.
The X post of Tibo did give us a chuckle thanks to Meguna (nudevise).
Meta has been on a roll this week, with the launch of "the world's first Personal AI agent" called Muse, powered by Muse Spark 1.3. Compared to a regular LLM, Muse actually takes on work, allowing for things like automated shopping, sending emails and booking travel. One of the main selling points for Muse is the way the model was constructed to be broadly usable. There is no need to have a PhD in machine learning, rather it works straight out of the box and builds context around the user along the way. Another interesting point is the security aspect of Muse. Meta specifically designed the system to run independently in the cloud and allows users to fully control what apps and services Muse connects to. The initial rollout is currently limited to adults in the US, so we'll have to wait to get our hands on it in Europe, but things are looking promising. Are Personal Agents the future? Perhaps. Are they cool as heck? Definitely. Will we let Muse shop for those new nikes? YES!
On a less upbeat note, Jacob Coxon announced that he resigned from Anthropic in a rather conspicuous post. The AI researcher actively warns against the power of AI technology, and specifically points the finger to OpenAI and Anthropic, describing them as "not acting responsibly". Furthermore, Coxon believes AI could kill us "by the end of the decade". Chilling thoughts but definitely not something to brush off as wackadoodle. Given that within this last month alone, many AI driven security exploits came to light, the safety debate is more relevant than ever. And Coxon is not alone in writing it on the wall, getting support from researchers across the field, including Antropic's Evan Hubinger. As this debate grows, we, and the remainder of the field, are continuously working on defining safety guidelines in all AI related products. And we urge enthusiasts of all kinds to do the same.
Google pledged a $15B/€13B investment in AI infrastructure in Finland. This marks the second largest investment in Europe from Google. Finland is currently booming with AI datacenters, due to the country's abundant and low-cost clean energy supply. With this investment, Google expands its regional hub, adding 3 new datacenters in northern Finland. On top of this, the deal constitutes a 22-year power purchase agreement with Fortum, guaranteeing continuous carbon-free electricity tied to extending the Loviisa nuclear power Plant. All in all, this investment marks a positive economic incentive in Europe and potentially opens up the gate for more companies to invest in Europe. In our opinion, Europe is falling behind compared to the US and China in terms of AI, and these types of investments are welcome.
As always, a lot more happened than we can comfortably condense into 5 minutes, so feel free to hit us up if you feel like we've missed something! For now, we wish you a good weekend and we'll be back next week!