OpenAI Struggles to Align Itself + Google Buried the Headline

It’s Hard Enough Aligning Humans

What Happened? Ilya Sutskever leaves OpenAI but despite firing Sam Altman personally, declares that he is ‘confident that OpenAI will build AGI that is both safe and beneficial under [his] leadership’. Jan Leike, co-head of Superalignment - the team that had just 3 years left on the clock to ‘align superintelligent AI’, declares his resignation in two words:

• Promised 20% of OpenAI’s recent total compute, the Superalignment team, in the 10 months since it’s formation, has published sparse new research but has splintered spectacularly, a foreboding sign for the effort.

• The team has no new leader in place, as of the 16th May, raising questions over OpenAI’s ongoing commitment to the initiative.

• Another OpenAI AGI Safety Researcher, Daniel Kokotajo, left this month - 'sacrificing 85% of his family’s net worth’ - to argue that OpenAI is not treating the safe arrival of AGI seriously enough. Another face of the team, Leopold Aschenbrenner, was fired for alleged leaking.

So What? One could frame this story as revenge for the aborted firing of Sam Altman by Sutskever. But the deeper significance is over the simpler story; it will be incredibly hard for humans to agree on what principles to embed within a superintelligent machine, let alone a set of people to trust with the responsibility. One can only hope we will rise to the occasion as the day where we will need to actually get this right draws closer.

Does It Change Everything? Rating = ⚂

Google Hid Amazing Demos Behind Those 123 ‘AI’ Mentions

What Happened? The Google I/O. If all of the products and demos contained within the almost 2-hour event had been announced just 100 days earlier, they would have shocked the world, but for real: 1080p AI video generation, 2m token context window [all of the Harry Potter series in one prompt], and the live, low-latency video conversation with Gemini as you roam a room [new demo]. But 3 months in AI news is a long time.

• Most impressive was the Project Astra video demo, that replicates much of the magic of GPT-4o. Hold your camera up to anything and simply ask questions, in a real time.

• We also got Gemini Flash, a small, fast, cheap, not-just-text model, again benefiting from a 1.5 million+ word context limit, and Veo, a Sora-like video generation model.

• But the event felt too long, and almost unfairly overshadowed by GPT-4o. It was also hampered by distant promises of releases ‘later this year’ and waitlists upon waitlists. Also, no mention of 1.5 Ultra or Gemini 2, models that might have rivalled - or exceeded - GPT-4o and Claude 3 Opus in intelligence.
  

So What? While Google may still end up ‘winning the race’ to AGI, and continues to print money with Search, they remain, at least for now, a step behind OpenAI in the realm of LLMs. Of course, zooming out a bit, they continue to crush in real-world utility with tools like AlphaFold 3.

Does It Change Everything? Rating = ⚀

Become a paying subscriber of Premium to get access to this post and other subscriber-only content.

Already a paying subscriber? Sign In.

A subscription gets you:

• • Exclusive posts, with hype-free analysis.
• • Sample Insider videos, hosted ad-free on YT, of the quality you have come to expect.
• • Access to an experimental SmartGPT 2.0 - see multiple answers to the same prompt w/ GPT-4 Turbo, for example, then have Claude 3 Opus review its work. Community-driven - so you can take the lead.
• • Support for balanced, nuanced AI commentary, with no middleman fees, for you or me. Would love one day to be in a position to have a small team of equally engaged independent researchers/journalists.