AI Leadership Shuffle: OpenAI, Microsoft, and the Future of AI Innovation

What happened? Over the last three days, Sam Altman, the CEO of OpenAI was fired, and Greg Brockman, their President and board chair, as well as head of research quit in response. Subsequently, OpenAI was in discussions to hire them both back, but Microsoft swooped in and convinced both Altman and Brockman to join Microsoft to build the future of AI (fun fact: Microsoft, a major investor in OpenAI, was just as caught off-guard by Altman’s firing as the rest of us).With their transition announced - key talent within OpenAI quickly and publicly started indicating that they will follow their CEO to Microsoft – all done and dusted between the close of business Friday and start of business Monday.  Wow.

This corporate version of “leadership musical chairs” has put the whole industry on edge.  While we never expected the rapid change to come like this – at Attainable Edge we’re firmly in the camp that the pace of AI technology acceleration is too rapid to lock ourselves into any one provider. So, we’ve built our solutions intentionally to be able to quickly adjust to competing model providers depending on the ever-changing price/performance balance.

Why does it matter? There is in fact no performance balance right now. When model performance and/or intelligence matters most, GPT4 has a commanding lead vs all competitors in the LLM space. One of the standard ways to compare the performance of large language models (LLMs) is a metric called Massive Multitask Language Understanding (MMLU) which measures how much knowledge is acquired into the model during pre-training by asking a standard set of questions –think of it as an SAT score built to assess AI performance. The test is comprised of 14,079 questions across 57 subject areas – here are two example questions from the MMLU (answers in italics):

• Philosophy - Nagel claims that most skeptical arguments:
  A. are the result of applying arbitrarily stringent standards.
  B. are based on linguistic confusions.
  C. are logically self-refuting.
  D. grow from the consistent application of ordinary standards.
  ‍
• Astronomy - What is true for a type-Ia supernova?
  A. This type occurs in binary systems.
  B. This type occurs in young galaxies.
  C. This type produces gamma-ray bursts.
  D. This type produces high amounts of X-rays.
  

Ok, so now we know how we can compare AI performance, how do the best LLMs compare?

While Open AI’s current lead is significant, it won’t remain so without additional work and investment. One of the most interesting things happening in the LLM space is what’s called the “Scaling Law” sometimes referred to as the “Theory of Intelligence” – research is currently showing that performance improves when we increase any one of three variables in model training (https://lastweekin.ai/p/the-ai-scaling-hypothesis). In simple terms – as we throw more resources at the current AI technology –it gets better and better with no theoretical limit in sight:

1) Training iterations / compute investment
2) Additional training data
3) Larger / more complex model

So what? What this means in practice – is the AI model competition is a money = performance game. It’s literally pay-to-win. Even Ilya Sutskever, Chief Scientist of OpenAI, has made the comment that the technical problems with scaling AI at this point are smaller than the problems of acquiring enough resources to implement the training.  GPT 3.5 cost roughly $10M to train, GPT4 cost$100M … models being trained now are estimated to be $1B to train – and it makes sense based on the above scaling laws to expect another tectonic leap in performance like we saw in the last generation (GPT3 -> GPT4).

To bring it back to current events – how does this performance scoring affect our response to the leadership changes in AI this weekend? From our viewpoint, Microsoft has put themselves in a win-win situation:

• Win #1 - If OpenAI maintains their lead under new leadership then Microsoft wins also as they maintain long-term licensing rights to use Open AI’s latest and greatest models up until OpenAI declares AGI based on Microsoft’s $13B investment into OpenAI. We view this outcome as slightly less likely given OpenAI’s new focus on slowing during to ensure AI safety, reinforced by the appointment of Emmet Shear as interim CEO, a advocate for AI risk management.
  ‍
• Win #2 – If a talent exodus develops where OpenAI is hamstrung by folks leaving to follow Sam and Greg to Microsoft, then Microsoft wins even more – by having the top talent in house and proprietary control of the top model. We currently see this as the most likely outcome.
  

What to expect: Considering those scaling laws of intelligence – the only folks with compute capabilities large enough to push forward the frontier of what’s possible are Google, Amazon, and Microsoft. Amazon is way behind; Google is making huge investments into their Gemini model, but they’re still playing catchup. The biggest risk to Microsoft would be from top talent leaving OpenAI to join Google or Amazon, but kudos to them for their swift hiring this weekend of top OpenAI talent before the rest of the industry even had a chance to respond. Based on their rapid response, we believe Microsoft has significantly mitigated the risk to the dominance of the OpenAI/Microsoft partnership.

Given our view of how this shakes out, we believe that closer integration with Microsoft is a safe bet for enterprise AI applications. Attainable Edge believes this weekend’s events will position Microsoft very well to win in the commercial market for at least the next 12-24 months – an eternity in this blazingly fast-changing industry. Given recent events and the current reality, Attainable Edge will be placing a few more bets on Microsoft’s horse.