Are We Back to Before? OpenAI 2.0, Inflection-2 and a Major AI Cancer Breakthrough
00:00:00.000 | OpenAI are back, but that doesn't mean nothing has changed.
00:00:03.720 | This video is about giving you more context, not only on what happened,
00:00:07.840 | but what it means and what else has been going on with AI
00:00:11.600 | that's arguably more important.
00:00:13.320 | And yes, it seems Sam Altman is also back at the helm of OpenAI,
00:00:17.920 | although interestingly, he said, I'm looking forward to returning to OpenAI
00:00:22.200 | this morning.
00:00:22.760 | Seems like the terms of that arrangement aren't yet finalized.
00:00:26.040 | One thing that's definitely not back in fashion
00:00:28.480 | and has arguably been lost in the chaos is the original OpenAI charter.
00:00:33.600 | It said it would be wise to view any investment in OpenAI Global
00:00:37.800 | in the spirit of a donation, with the understanding that it may be difficult
00:00:41.480 | to know what role money will play in a post-AGI world.
00:00:45.240 | Furthermore, the charter takes precedence over any obligation
00:00:49.040 | to generate a profit.
00:00:50.280 | The company may never make a profit and the company is under no obligation
00:00:54.600 | to do so.
00:00:55.200 | With a tentative $90 billion valuation,
00:00:58.280 | I wonder how many at OpenAI would sign up to that charter again now.
00:01:02.520 | Anyway, back to Las Vegas, where Sam Altman was attending
00:01:05.480 | the biggest party of the weekend, the Formula One race,
00:01:08.640 | when he was fired by the board.
00:01:10.440 | Many of the possible reasons for that were covered in my previous video,
00:01:14.040 | but Ilya Sutskevo was asked directly to provide a reason.
00:01:17.560 | An hour and a half after firing Sam Altman, Ilya Sutskevo was asked
00:01:21.240 | by OpenAI employees, will they ever find out what Sam Altman did?
00:01:25.800 | To which he replied, no.
00:01:27.680 | This is according to the Wall Street Journal.
00:01:30.240 | That reluctance apparently went so deep that even when the new CEO,
00:01:34.200 | Emmet Scheer, asked the board, demanded the board to provide concrete evidence
00:01:38.760 | of Sam Altman's wrongdoing, they did not do so.
00:01:41.640 | He then said he could not remain in the CEO role and tried to help
00:01:45.800 | both sides find common ground to get Sam Altman reinstated.
00:01:49.600 | Speaking of the outgoing CEO, Emmet Scheer, apparently when he announced
00:01:53.160 | an all hands meeting, fewer than a dozen people showed up
00:01:56.520 | and apparently several employees responded with middle finger emojis.
00:02:00.560 | That follows on from the executive team of OpenAI pressing the board
00:02:05.080 | for 40 minutes for specific examples of Sam Altman's lack of candor or honesty.
00:02:10.520 | The board refused, apparently citing legal reasons.
00:02:13.440 | And it seems at one point, law enforcement entities such as the US
00:02:17.440 | Attorney's Office even got involved.
00:02:19.480 | At this point, it seems a bit more of the reasoning came out.
00:02:22.800 | The board intimated that there wasn't one incident that led
00:02:26.160 | to their decision to eject Altman, but instead a consistent,
00:02:29.400 | slow erosion of trust over time that made them increasingly uneasy.
00:02:33.640 | And as I mentioned in my previous video, also Sam Altman's mounting list
00:02:37.400 | of outside AI related ventures, like creating an AI chip,
00:02:41.400 | which raised questions for the board about how OpenAI's technology
00:02:44.760 | or intellectual property could be used.
00:02:46.800 | The board then discussed the matter with their own legal counsel.
00:02:50.160 | And after a few hours, they returned, still unwilling to provide specifics.
00:02:54.440 | They said that Sam Altman wasn't candid and often got his way.
00:02:57.960 | And they said that he'd been so deft that they couldn't even give a specific example.
00:03:03.640 | Indeed, the only person to go on record about any lying actually works at Google DeepMind.
00:03:08.400 | This is Jeffrey Irving, who used to work at OpenAI.
00:03:11.840 | He said Sam Altman lied to him on various occasions, was deceptive, manipulative and worse to others.
00:03:17.360 | He said that when he lied to him, Jeffrey Irving,
00:03:20.120 | it was about other people to try to drive wedges.
00:03:23.040 | But according to the New York Times, the firing might not have even been about lying after all.
00:03:27.680 | Apparently, Sam Altman met one of the board members, Miss Toner, to discuss a paper she had written.
00:03:33.040 | Sam Altman then wrote an email to his colleagues reprimanding Miss Toner for that paper,
00:03:38.000 | saying it was dangerous to the company, particularly at a time when the FTC was investigating OpenAI.
00:03:43.400 | Miss Toner defended her paper as an academic paper, which I have read in part and I'll get to in a moment,
00:03:48.280 | and said it was analyzing the challenges that the public faces when trying to understand the intentions of the companies developing AI.
00:03:55.280 | But Sam Altman disagreed and said, "I did not feel we're on the same page on the damage of all of this.
00:04:00.760 | Any amount of criticism from a board member carries a lot of weight.
00:04:04.120 | Honestly, though, before today, I hadn't heard of the paper or heard anyone discussing it."
00:04:08.560 | So what does the paper say?
00:04:10.080 | Well, on page 30, we have this.
00:04:12.120 | First, heaps of praise for OpenAI for publishing a system card outlining the risks of GPT-4.
00:04:19.360 | After that praise, though, they do say that the release of ChatGPT four months earlier overshadowed the import of that system card.
00:04:27.920 | The paper then describes how ChatGPT's release sparked a sense of urgency inside major tech companies.
00:04:34.160 | One thing the paper did do, though, was praise Anthropic, the rival to OpenAI.
00:04:40.240 | It singled out their different approach of delaying model releases until the industry had caught up.
00:04:46.720 | It quoted how Anthropic's chatbot, Claude, had been deliberately delayed in order to avoid advancing the rate of AI capabilities progress.
00:04:55.360 | Now, many people won't remember this, but this was actually cited in the GPT-4 technical report.
00:05:00.760 | That report, released in March, warned of acceleration as a risk to be addressed,
00:05:06.840 | and advisors that they employed told them to delay GPT-4 by a further six months and be a bit quieter in the communication strategy around deployment.
00:05:16.480 | Toner's paper did then describe the, quote, "frantic corner cutting" that the release of ChatGPT appeared to spur.
00:05:24.560 | But honestly, it's a fairly nuanced report and then critiques Anthropic's approach, saying it's ignored by many.
00:05:31.480 | The report says, "By burying the explanation of Claude's delayed release in the middle of a long, detailed document posted to the company's website,
00:05:39.640 | Anthropic appears to have ensured that this signal of its intentions around AI safety has gone largely unnoticed."
00:05:45.960 | And before you think that she must be alone at OpenAI in critiquing the release of ChatGPT, Sam Altman himself listed it as a possible future regret.
00:05:55.960 | We're definitely going to have some huge regrets 20 years from now. I hope what we can say is that we did far, far, far more good than bad.
00:06:04.160 | And I think we will. I think that's true. But the downside here is pretty big, and I think we feel that weight every day.
00:06:10.360 | Honestly, I think if we're going to regret something, it may be that we already pushed the button.
00:06:16.840 | Like, we've already launched this revolution. It's somewhat out of our hands.
00:06:21.000 | I think it's going to be great, but this is going to happen now, right? We're out of the... the world is out of the gates.
00:06:28.600 | I guess the thing that I lose the most sleep over is that we already have done something really bad.
00:06:34.680 | I don't think we have, but the hypothetical that we, by launching ChatGPT into the world, shot the industry out of a railgun,
00:06:44.040 | and we now don't get to have much impact anymore, and there's going to be an acceleration towards making these systems,
00:06:52.280 | which again, I think will be used for tremendous good, and I think we're going to address all the problems.
00:06:57.320 | But maybe there's something in there that was really hard and complicated in a way we didn't understand.
00:07:02.280 | And, you know, we've now already kicked this off.
00:07:04.520 | But if that's more of the context, what about what actually happened on the weekend?
00:07:08.640 | Well, Sam Altman's house became apparently a war room filled with OpenAI employees, including Mira Murati, the interim CEO.
00:07:17.240 | They began to use X or Twitter in a coordinated fashion for their campaign.
00:07:22.280 | This was also the time of the open letter to the board of directors from the employees of OpenAI.
00:07:29.120 | By my last count, it was signed by 747 out of the 770 OpenAI employees.
00:07:36.480 | It described how the board stated that allowing the company to be destroyed wouldn't necessarily be inconsistent with the mission of creating safe and beneficial AGI.
00:07:47.240 | The employees then threatened to resign from OpenAI and all join Microsoft.
00:07:52.920 | Notice how the threat is to take this step imminently unless all current board members resign.
00:07:59.200 | As of now, though, it seems like Sam Altman has agreed not to return to the board while Adam D'Angelo will stay.
00:08:05.560 | Furthermore, Sam Altman has agreed to allow an investigation into his affairs before the firing.
00:08:10.960 | Remember that Adam D'Angelo, the CEO of Quora and one of the people who voted to remove Sam Altman,
00:08:17.200 | was described by Sam Altman in a blog post from a few years ago as one of the few names that people consistently mention when discussing the smartest CEOs in Silicon Valley.
00:08:27.560 | And Sam Altman said he has a very long term focus, which has become a rare commodity in tech companies these days.
00:08:34.360 | I wonder if he would reaffirm that about Adam D'Angelo now.
00:08:37.920 | But back to Microsoft, who had not only announced the arrival of Sam Altman,
00:08:42.280 | but even gotten a floor of their office in San Francisco ready with laptops and clusters of GPUs.
00:08:48.840 | Satya Nadella, the CEO of Microsoft, even said that Sam Altman and Greg Brockman were in the process of joining Microsoft.
00:08:57.280 | So how are you envisioning this role with this sort of advanced AI team that Sam and Greg would be joining and leading?
00:09:04.560 | Can you explain that? And are they actually Microsoft employees right now?
00:09:09.280 | Like, who do they work for?
00:09:11.600 | Yeah, so they're all in the process of joining.
00:09:14.480 | And yes, I mean, the thing is, look, we have a ton of AI expertise in this company.
00:09:19.840 | And Kevin Scott, the CTO of Microsoft, said this.
00:09:22.800 | We have seen your petition and appreciate your desire potentially to join Sam Altman at Microsoft's new AI research lab.
00:09:30.080 | Speaking as if Sam Altman had already joined.
00:09:32.320 | Know that if needed, you have a role at Microsoft that matches your compensation and advances our collective mission.
00:09:38.160 | So what does Microsoft now make of OpenAI not joining them and instead staying independent?
00:09:44.080 | Satya Nadella said this.
00:09:45.200 | We are encouraged by the changes to the OpenAI board.
00:09:48.640 | We believe this is an essential first step on a path to a more stable, well-informed and effective governance.
00:09:55.520 | Sam, Greg and I have talked and agreed they have a key role to play.
00:09:59.680 | Notice it's not really defined with the OpenAI leadership team in ensuring OpenAI continues to thrive and build on its mission.
00:10:06.640 | But a more revealing quote from Nadella came last night in an interview with Kara Swisher.
00:10:11.280 | We're never going to get back into a situation where we're surprised like this again.
00:10:16.400 | And Bloomberg are reporting that the final reworked OpenAI board will have up to nine new directors,
00:10:23.280 | with Microsoft getting possibly one or more seats.
00:10:26.960 | One of the new board members will be Larry Summers, and I'll get to him in a moment.
00:10:31.440 | But first, there's something that slipped out during the chaos that I want to pick up on.
00:10:35.840 | At one point during the weekend, OpenAI's board approached Anthropic about a merger.
00:10:41.440 | And I'll give myself a little bit of credit for predicting that in one of my previous videos.
00:10:46.560 | Anyway, apparently that was quickly turned down by Dario Amadei, the CEO of Anthropic.
00:10:51.920 | I'm going to get to other AI news in a second, but first a quick look at the views of Larry Summers on AI.
00:10:58.480 | He views language models as a major wake-up call for the white-collar worker.
00:11:03.520 | I think it's coming for the cognitive class.
00:11:07.520 | Chet CPT is going to replace what doctors do.
00:11:11.280 | Hearing symptoms and making diagnoses before it changes what nurses do.
00:11:18.000 | Before we leave this OpenAI saga, there's one more thing I wanted to point out.
00:11:22.320 | I'm going all the way back to that OpenAI charter that I introduced in the beginning.
00:11:26.560 | And the fifth clause says this, "The board determines when we've reached AGI."
00:11:31.600 | Again, by AGI, we mean a highly autonomous system
00:11:35.280 | that outperforms humans at most economically valuable work.
00:11:38.880 | When the board determines that we have attained AGI,
00:11:42.160 | such a system is excluded from IP licenses and other commercial terms with Microsoft.
00:11:47.520 | But unless OpenAI and the board want this entire saga to repeat again,
00:11:52.720 | I think they really need to better define AGI.
00:11:55.920 | What does highly autonomous mean?
00:11:57.920 | More importantly, what does outperforming humans at most economically valuable work even mean?
00:12:02.720 | The AI and broader technology we have today, if we transported it back 200 years in the past,
00:12:08.560 | would be able to outperform most humans at most economically valuable work.
00:12:13.200 | Obviously, what happened is that humans reacted to that and changed how they work.
00:12:17.360 | So this debate could arise again in the not too distant future.
00:12:20.720 | OpenAI talk about aligning superintelligence within four years.
00:12:24.240 | But if the board declares AGI, Microsoft might be able to say,
00:12:27.760 | "Well, we haven't yet reached post-labor economics.
00:12:30.320 | People are still contributing in economically valuable ways as carers,
00:12:34.800 | or as entertainers, or YouTube hosts, or whatever else happens."
00:12:38.560 | I do feel the world needs to come up with a much more rigorous definition of AGI.
00:12:43.920 | So yes, I agree with Mustafa Suleiman that it has been an utterly insane weekend.
00:12:48.960 | But it's time at last to move on to other AI news.
00:12:52.960 | Because apparently, Inflection AI have finished training
00:12:56.000 | the second best large language model in the world.
00:12:59.440 | Some will be surprised that it's only the second best LLM, still behind GPT-4.
00:13:04.160 | It does though say that we're scaling much further.
00:13:07.680 | And they are more than capable of doing so.
00:13:10.000 | To train Inflection 2, the startup used only 5,000 NVIDIA H100 GPUs
00:13:17.280 | when they have a supercomputer with 22,000 H100s.
00:13:22.240 | Nevertheless, you can see the improvements from Inflection 1 to Inflection 2
00:13:27.440 | beating Palm 2, which now powers BARD.
00:13:30.400 | It also apparently beats XAI's Grok 1 and Anthropic's Claude 2.0 model.
00:13:35.760 | And Anthropic, who we've already mentioned multiple times today,
00:13:39.120 | have introduced Claude 2.1.
00:13:41.520 | Notice the number there, not going for the dramatic 2.5 or 3, just 2.1.
00:13:47.040 | That's kind of an admission that this is quite a small iterative update.
00:13:51.520 | Essentially, they've scaled up their context window to 200,000 tokens.
00:13:55.920 | That's maybe 150,000 words.
00:13:58.400 | If this was open AI, this would probably be an entire event,
00:14:01.520 | but it's just a blog post.
00:14:03.040 | Apparently, they've made it so that Claude 2.1 is much more honest
00:14:07.120 | about when it doesn't know the answer to a hard question.
00:14:09.840 | Instead of just blurting out a hallucination,
00:14:12.400 | it will be much more likely just to say, "I don't know."
00:14:15.040 | The number of errors that Claude 2.1 makes about questions relating to the middle,
00:14:19.840 | beginning and end of long documents seems to have gone down.
00:14:24.160 | But that definitely doesn't mean that it's reliable for questions,
00:14:27.600 | particularly if those questions pertain to the middle of the document.
00:14:31.680 | Notice if the fact that you're trying to find comes right at the beginning,
00:14:35.120 | this top row, or right at the end of a document, it will do well.
00:14:39.280 | But especially when you're dealing with those really long books
00:14:43.120 | or multiple documents up to 200,000 tokens,
00:14:46.240 | the errors for the middle can be quite dramatic.
00:14:48.800 | Compare that to a shorter, say, 15,000 word document.
00:14:52.800 | This column is all green, even for the middle of that document.
00:14:56.960 | But all of what I've covered in this video pales in comparison
00:15:00.480 | to AI advances like this one.
00:15:02.720 | This was published in Nature on Monday.
00:15:05.280 | AI detection of a certain type of pancreatic cancer,
00:15:08.800 | pancreatic ductal adenocarcinoma, apparently the most deadly solid malignancy,
00:15:14.400 | has now overtaken mean radiologist performance,
00:15:17.760 | translated AI doing better than human doctor performance.
00:15:21.840 | And by a fair margin, 34% in sensitivity and 6% in specificity.
00:15:26.640 | Sensitivity is about not missing cancers that are there,
00:15:29.840 | and specificity is about not saying there is a cancer when there isn't.
00:15:33.680 | But in both cases, it's outperforming mean radiologist performance.
00:15:37.520 | This is an AI advance that could genuinely save thousands of lives.
00:15:41.840 | And on that crazy optimistic note, I want to end the video.
00:15:45.120 | Thank you as always for watching to the end.
00:15:47.680 | I promise to get to more non-open AI AI news in the future,
00:15:52.000 | and have a wonderful day.