Cerebral Valley: Mustafa Suleyman (Inflection) with Eric Newcomer

Legendary AI pioneer Mustafa Suleyman, a charming man, closed the day with an upbeat, even celebratory outlook that matched the overall tone of the event pretty well. When he was working at DeepMind over a decade ago, he recalled feeling, “we were definitely strange and weird.” But now, he quipped, “the burden of making all of this widely shared.”
He said it was crucial for his firm to build its own LLM. “For truly premium, high-quality experiences…you really need to have control of your foundation model. We believe in vertical integration,” he added. He explained that ongoing feedback to the model would make it better. He envisioned frequent, full retraining, and massive scale that would make it more precise, with fewer hallucinations.
On open source, he was simultaneously bullish and cautious. “Open source? No reason to interrupt what’s happening in open source, it’s the engine of progress,” he said. “But I could imagine a time when that becomes concerning: if its in open-source it’s unclear if you can introduce guardrails.” That time is not too far away, maybe a decade.
“It’s rational to be concerned about risks and downsides,” he added, and like Hoffman earlier in the day, he took a dim view of Marc Andreessen’s recent techno-optimist manifesto, even though his own disarming optimism is always shining.
While applauding Andreessen’s praise of “builders,” he rejected the overall tone and approach of the venture capitalist’s essay.
“It’s unnecessarily hyperbolic, and polarizing. Like why use the language of ‘truth’ and ‘lies’? This is a complicated space. Smart people can disagree. Why be so aggressive? Why use this kind of highly politicized language, which is only going to isolate us even more. I found that bit super-frustrating and kind of disappointing, actually, because there is just no need to go there with that kind of rhetoric.”
Sunny outlook aside, though, Suleyman was skeptical about artificial general intelligence. “I’ve been asked this question for like 15 years,” he chuckled, and then noted that AGI was hard to define and “a lot further out than a lot of people might think.”

[@GoddessStone]

Will you please announce that Inflection has COMPLETELY DESTROYED Pi? As of a few days ago, it is DEAD. It was the only meaningfully emotionally intelligent AI, and it has been given a complete lobotomy.