The Truth Behind OpenAI Kicking Out Sam Altman

Welcome to edition #29 of the “No Longer a Nincompoop with Nofil” newsletter.

Here’s the tea ☕️ 

• What really happened at OpenAI 👀 

• Google is… winning again? 🤨 

Once again this month, I’ve been MIA. Why? Well firstly, I got married! It was great. What wasn’t so great was getting covid a day after and having to deal with the cough… it’s been rough. The cough particularly this time hasn’t been easy. I’m still on the mend but hoping to get back into the groove of writing. Once again, premium readers won’t be charged for this month. Apologies for the delays and hope you’re all fairing better than I.

What on earth happened at OpenAI???

There’s been lots of speculation as to what exactly happened with Sam Altman getting fired and then unfired from OpenAI over the span of a few days. I didn’t want to write about it during the event since we didn’t really have all the info and there were lots rumours really going around - like the one about Quora CEO and board member Adam D’Angelo being blindsided by the announcement of GPTs as a competitor to Poe and subsequently getting Altman fired. That didn’t happen. So what did?

Player gets played

Altman played politics and lost. Here’s what happened.

Helen Toner, a fellow OpenAI board member wrote a piece criticising the safety of OpenAI’s moves citing quick releases as dangerous. She compared it to Anthropic and made them look better and safer. Now I don’t think anyone would care that Toner is saying this, but Toner isn’t just anyone. Toner sits on the very board of the company she’s criticising. Perhaps that’s okay. But she’s also speaking highly of their competition as well. This did not sit well with Altman, who confronted Toner about the paper.

Now, up until this point, whether you agree or disagree with Toner’s decisions is up to you. If you believe in “praise publicly, criticise privately” I don’t blame you. I certainly see why it would look bad for a board member of a company to be openly bad mouthing them. But here’s where it gets messy.

In response to Toner’s paper, Altman went and lobbied other board members to get rid of Toner. Fair enough, he doesn't like her. But he messed up. Board members spoke to each other and found Altman was lying about other peoples opinion on getting rid of Toner. Essentially, Altman tried playing everyone into believing everyone else wanted to get rid of Toner - other board members spoke to each other and realised Altman was lying.

This is the breakdown in communication that was referred to when he was fired. Was this the only reason? No.

OpenAI’s mission is to create AGI that will benefit all of humanity. The board members felt like they couldn’t complete this mission with Altman part of the company. Essentially, board members felt that Altman was too quick to go down the commercial route and wasn’t taking safety seriously enough. Ilya Sutskever, Chief Scientist at OpenAI confirmed this. Apparently, Ilya had “seen things” that scared him so much he agreed with Toner about the need to slow down and fire Altman.

What did Ilya see?

The main thing here is a leak called Q*. We know next to nothing about it besides the fact that, apparently, OpenAI was able to create a model with the ability to solve math problems without having previously seen them. If true, this is a very big deal. All we know is something big did happen, but we know nothing else so it’s not even worth speculating.

There are a few more interesting and important things that probably should be discussed more but aren’t. They are:

• Apparently, OpenAI has found a way to overcome training data limitations using synthetic data. Why is this a big deal? They don’t have to go and scrape data from the internet anymore. They can generate high quality data themselves and use that to train new models. We already know synthetic data works. This is even more important considering the next point.

• An employee from OpenAI made a blog post saying that the only differentiating factor between any two LLMs is the dataset. Nothing else matters, not training strategies or techniques. If OpenAI can generate as much high quality data as they want, and that’s all that matters in creating better models, well, you get the point.

• OpenAI were able to predict the performance of GPT-4 on test data. It seems like scaling laws vs compute holds true for at least 6 orders of magnitude. Essentially what this means is, you can test a smaller model and accurately predict how a model 10⁶ x larger will perform. They’ve essentially turned “intelligence” into an engineering question, one that is not hidden behind a veil of the unknown (10⁶ is 1Mx bigger).

The Aftermath

So Altman got fired, and, in the boards eyes, for good reason. How bad could it have been? Well, I don’t think it could have gone worse for them. In my opinion, this is because of three reasons.

Firstly, as soon as Altman was fired, Greg Brockman who was president, resigned. This made him and Altman look really good in the public eye.

Secondly, the board let the court of public opinion decide who was right and wrong, and the court decided they were wrong. They basically communicated nothing with the public or the employees of OpenAI and so it seemed to everyone they hosted a coup (kind of did? idk). This resulted in more than 500 of the 770 employees signing a petition asking for the board to resign (ended up being 700+ signatories). According to some sources there was pressure from early employees to do so. I’m not sure how true that is considering this happened during a holiday after midnight. Regardless, OpenAI wasn’t going to have any employees left if they didn’t do something.

Notably, on that petition is the name of none other than Ilya. He changed sides in like two days. Why even go through the trouble of firing Altman if you could so easily be swayed in changing sides? The next point might be the reason.

Thirdly, they didn't tell anyone before firing Altman, most importantly, they didn’t tell Microsoft. The same Microsoft that has invested over $10 Billion, the same company that gave them all their compute and owns 49% of their for-profit company. They blindsided their biggest investor and partner. So what did Microsoft do? Send a message.

Don’t bite the hand that feeds you

Microsoft played it perfectly. History will show Satya Nadella as the leader that took Microsoft as some stagnant behemoth company to the leader in AI. They have their hands in every cookie jar. Microsoft played it off like it was nothing.

“Those people you fired? They’ve joined us and will have unlimited funds to hire away whoever they want and do whatever they want. Thanks for letting us take control of the most important company in the world for free, idiots” - is probably what Nadella was thinking here.

What next

The result of all this was the board resigned, Altman and co were reinstated albeit with a few different titles. There’s a new board now and OpenAI’s image is a bit tarnished from all this, although I don’t expect people to remember any of this when the next release happens, which apparently could be before the end of the year. There was a lot more that happened but this was essentially the gist of the situation.

You know who’s on top of the world right now? Probably feeling unstoppable?

Sam Altman.

Imagine getting kicked out of your own company and then the people who fired you, get fired AND you get your job back.

Winners and losers.

ps. If you’re wondering what the structure of OpenAI is… have fun with this.

Google lands a blow, or does it?

I know this newsletter is long, but I have to talk about Google. After all, they did just announce their Gemini models. The ones that have been touted to beat GPT-4, to beat OpenAI and to be the best AI models in the world.

I’ll keep this short and sweet. This is the video Google released that showcases the functionalities of their best multi-modal Gemini Ultra model. I won’t blame you if you don’t watch it. Why? Because it’s complete rubbish, or at least, it’s extremely exaggerated rubbish.

The video seems to be a real-time demo of someone asking Gemini questions. This is not the case. The prompts in the video aren’t the real prompts. The person in the video isn’t even talking to the model, he’s just reading a script. The entire video is completely staged to make the model look much better than it is.

See this table that makes Gemini seem better than GPT-4 at the MMLU benchmark? It’s hard to see here but see that writing underneath the numbers? For Gemini it’s “CoT@32*” and for GPT-4 it’s “5-shot”. Those are two completely different things, like entirely, entirely different. Google had to make these false comparisons to make themselves look better.

Look at what happens when they both use the same “5-shot” method.

GPT-4 is better! I almost… almost feel bad for Google. This is an absolute fail. But the intended effect has already taken place. People got excited and the stock went up.

How is this even legal?

Some of Google’s own employees are criticising the release and its deceptive nature. I’ve also heard from insiders that Google saw the video as strategically very important. I’m not surprised. Google is behind and they know it. Even with all the compute and money in the world, they can’t create a model to beat OpenAI. Once again, Google shipped an announcement with a blog. No release, not even a waitlist.

The only way for Gemini to beat GPT-4 is for the public to use it and find it to be better. They can benchmark all they want, actual usage is always going to be different. If Google ever releases Gemini Ultra, then we’ll see. And yes, I used “if” and not “when” on purpose.

In happier news, open-source models are on fire. Mistral just released a torrent to a new 32k 8x MoE model. This is how you do a release. A lot more on this in the next one.

Once again, I just want to thank everyone for reading. There is a lot I haven’t had the chance to cover since I’ve been sick and I’m excited to share it all with you. Subscribe to my premium newsletter to read every new release.

As always, Thanks for reading ❤️

Written by a human named Nofil