back to index

9 New Gemini Leaks, Code Llama and A Major AI Consciousness Paper


Whisper Transcript | Transcript Only Page

00:00:00.000 | Like buses, AI news can sometimes be slow and sometimes arrive all at once.
00:00:06.020 | In the last few days we have had dramatic new leaked insights into the sheer breadth
00:00:11.680 | of Google's Gemini. Just today we've had the release of Meta's Code Llama and earlier their
00:00:18.080 | impressive multilingual seamless M4T model. And last but definitely not least, this 88 page
00:00:26.240 | AI consciousness report. And yes, I read it all, it's juicy so I'm saving that for the end.
00:00:32.640 | But let's start with two major paywalled articles, one from the Information and one from the New
00:00:38.620 | York Times, about Google's Gemini model. From both of them I counted a total of 9 new revelations,
00:00:45.000 | so let's get straight to it. To give you a sense of timeline by the way, Google's newly merged AI
00:00:50.360 | SWOT team they call it, is preparing for a big fall or autumn launch. The takeaway
00:00:56.040 | for me from both articles is that Gemini is going to be the everything model. Did you
00:01:01.360 | know it's going to be the rival to Midjourney and Stable Diffusion? Midjourney only has
00:01:07.240 | 11 full time staff so it is more than plausible that Google's Gemini could outperform Midjourney
00:01:13.960 | version 5. Next we may be able to create graphics with just text descriptions and control software
00:01:20.940 | using only text or voice commands. These next two are speculation so I'm not even
00:01:25.840 | counting them in the list of leaks. But I've already covered in a previous video that Gemini
00:01:30.660 | has been trained on YouTube video transcripts. And the speculation is that by integrating video
00:01:36.780 | and audio into Gemini, it could perhaps help a mechanic diagnose a problem with a car repair
00:01:42.880 | based on a video. Or be a rival to Runway ML by generating advanced text to video based on
00:01:49.340 | descriptions of what a user wants to see. You can start to see why I'm beginning to think of it as
00:01:54.080 | the everything model.
00:01:55.640 | Another leak is that one of the co-founders of Google, Sergey Brin is working on the front
00:02:00.520 | lines of Google Gemini. And lastly from this article I found it really interesting that
00:02:05.860 | Google's lawyers have been closely evaluating the training and they made researchers remove
00:02:11.640 | training data that had come from textbooks. Even though those textbooks helped the model
00:02:17.040 | answer questions about subjects like astronomy or biology. And I do wonder if they privately
00:02:22.360 | benchmarked Gemini before removing that crucial data.
00:02:25.440 | But if that's not enough, prepare to also receive life advice. My theory here is that Google wants to compete
00:02:33.000 | directly for market share with Inflection's pie. What if you want scientific, creative or professional
00:02:39.780 | writing? Yep, they're working on that too. In fact, we already know that Google has software
00:02:44.500 | named Genesis that they're pitching to the New York Times, which can generate news articles,
00:02:49.460 | rewrite them, suggest headlines, etc. But some people will be more interested in this feature that
00:02:55.240 | Google DeepMind is working on. The ability to draft critiques of an argument and generate quizzes,
00:03:00.780 | word and number puzzles. It's almost easier at this point to ask what might Google Gemini not
00:03:06.880 | be able to do. And yes, this is not Gemini, but Google DeepMind is also using AI to design the
00:03:13.200 | next generation of semiconductors. But if the fall seems far away, how about today when we got
00:03:19.260 | Code Llama from Meta? I spent much of the last two hours reading most of the 47 page
00:03:25.040 | paper and you can see Code Llama in action on screen. Some highlights include that the Code Llama
00:03:31.040 | models provide stable generations with up to 100,000 tokens of context. Obviously, that could
00:03:37.020 | be used for generating longer programs or providing the model with more context from your code base
00:03:42.240 | to make the generations more relevant. It comes in three versions, Code Llama, Code Llama Instruct,
00:03:47.580 | which can better understand natural language instructions, and Code Llama Python,
00:03:51.660 | better, of course, at Python. It's available for commercial
00:03:54.840 | use. And as you can see, some of the versions rival GPT 3.5 on human eval. That top score of
00:04:01.880 | 53.7% on Passat1 puts it in the same ballpark as Phi1. I've actually done a full video on Phi1,
00:04:09.340 | so do check that out. But that got 50.6%. But it is about 25 times smaller at 1.3 billion parameters.
00:04:16.860 | Interestingly, the Code Llama paper, which also came out about two hours ago,
00:04:20.980 | mentions Phi1 directly, saying that it follows in a similar,
00:04:24.800 | similar spirit, but the difference is that Phi1 is closed source.
00:04:28.680 | Anyway, a couple more interesting things before we move on from Code Llama. And the first one
00:04:34.140 | is the self-instruct method that they use. Let me know if you also find this fascinating,
00:04:39.080 | because step one was to generate 62,000 interview-style programming questions by
00:04:45.020 | prompting Llama2, the 70 billion parameter model. Then they removed duplicates in step two.
00:04:50.320 | But here's where it gets interesting. For each of those questions, they first generated a unit test,
00:04:54.760 | by prompting Code Llama 7 billion parameters. Then they generated 10 Python solutions by prompting
00:05:01.260 | Code Llama. Finally, they ran unit tests on those 10 solutions, and they added the first solution
00:05:06.800 | that passes those tests, along with the corresponding question and test, to the self-instruct
00:05:11.600 | dataset. If that sounded a bit complicated, let me try to distill it a bit. They asked the big brother
00:05:16.900 | Llama2 model to generate questions, then got the little brother Code Llama to generate tests for
00:05:23.680 | those questions. Then they added the little brother Code Llama to generate tests for those questions.
00:05:24.720 | Then got the model to generate solutions to its own tests, found the good solutions that don't
00:05:30.140 | forget it produced, and then used those to further train the model. To be honest, synthetic data and
00:05:35.860 | self-instruct seem to be the future of feedback. One final interesting quote from the paper on
00:05:42.360 | safety, and that was an argument advanced by one of their red teamers. They made the point that
00:05:47.600 | various scripts and code is readily available on mainstream public websites, hacking forums, or the
00:05:54.680 | web. And the advanced malware development is beyond the current capabilities of available LLMs. And even
00:06:00.640 | an advanced LLM paired with an expert malware developer is not particularly useful at the moment,
00:06:06.400 | as the barrier is not typically writing the malware code itself. Let me know what you think in the
00:06:12.100 | comments. But we must move on to Seamless M4T released a couple of days ago from Meta, which
00:06:18.980 | frankly seems amazing for multilingual translation. That speech to text, speech to text, speech to text
00:06:24.640 | to speech, text to text, and more. It has speech recognition for nearly 100 languages and can output in 36
00:06:32.640 | languages. But there's one feature I find particularly cool.
00:06:35.840 | Now, let's talk about code switching. Code switching happens when a multilingual speaker
00:06:43.340 | switches between languages while they're speaking. Our model Seamless M4T automatically
00:06:48.840 | recognizes and translates more than one language when mixed in the same sentence. As a multilingual
00:06:54.600 | speaker, this is a very exciting capability for me. I often switch from Hindi to Telugu when I speak
00:07:00.640 | with my dad. Notice in the following example when I change languages.
00:07:05.100 | I can speak Hindi, Telugu, and English. Sometimes, I can use all three languages in one sentence.
00:07:17.940 | I can speak Hindi, Telugu, and English. Sometimes, I use English. Sometimes, I use English. Sometimes, I
00:07:24.560 | use all three languages in one conversation.
00:07:27.860 | Speaking of cool though, we had this epic story out yesterday. AI gave a paralyzed woman her voice back.
00:07:35.260 | In a moment, you're going to see her being plugged in to the model. There we go. And the short version
00:07:41.900 | is that this woman suffered a stroke that left her unable to speak. But now, for the first time,
00:07:48.260 | her speech and facial expressions can be synthesized from her brain signals. Decoding these
00:07:54.520 | signals into text at nearly 80 words per minute, up from 14 words per minute.
00:07:59.860 | But let's now end on this, an 88-page report on consciousness in artificial intelligence, which counts as one of its
00:08:06.720 | co-authors, Yoshua Bengio, the Turing Award winner. It was dense and quite technical, but well worth the read.
00:08:14.060 | Look at this sentence in just the abstract.
00:08:16.680 | "Our analysis suggests that no current AI systems are conscious, but also suggests that there are no obvious technical barriers to build
00:08:24.480 | AI systems which satisfy these indicators."
00:08:28.020 | These are the indicators and each one gets a few pages in the report. And the reason that they're split up is because each one rests on a certain theory of consciousness.
00:08:39.020 | Obviously, the key problem is that we don't have a consensus theory on what consciousness is or how it comes about.
00:08:46.640 | So in a way, to hedge their bets, they group in different theories and look at the kind of indicators that would satisfy each one.
00:08:54.440 | Let's say that list seems so theoretical, why not just test the model or even ask the model?
00:08:59.680 | For more on that approach, see my theory of mind video.
00:09:02.940 | But the problem is, as they say on page four, the main alternative to a theory heavy approach is to use behavioral tests for consciousness.
00:09:12.440 | But as I talked about in the other video, that method is unreliable because AI systems can be trained, of course they are, to mimic human behaviors,
00:09:20.740 | are working actually in very different ways.
00:09:23.340 | Essentially, LLMs are the most efficient way to manipulate human behavior.
00:09:24.400 | They can be used to manipulate the behavior of other people.
00:09:25.940 | They can be used to manipulate other people's behavior.
00:09:27.240 | They can be used to manipulate other people's behavior.
00:09:29.240 | They can be used to manipulate other people's behavior.
00:09:31.240 | They can be used to manipulate other people's behavior.
00:09:33.240 | They can be used to manipulate other people's behavior.
00:09:35.240 | They can be used to manipulate other people's behavior.
00:09:37.240 | They can be used to manipulate other people's behavior.
00:09:39.240 | They can be used to manipulate other people's behavior.
00:09:41.240 | They can be used to manipulate other people's behavior.
00:09:43.240 | They can be used to manipulate other people's behavior.
00:09:45.240 | They can be used to manipulate other people's behavior.
00:09:47.240 | They can be used to manipulate other people's behavior.
00:09:49.240 | They can be used to manipulate other people's behavior.
00:09:51.240 | They can be used to manipulate other people's behavior.
00:09:53.240 | They can be used to manipulate other people's behavior.
00:09:54.360 | They can be used to manipulate other people's behavior.
00:09:56.360 | They can be used to manipulate other people's behavior.
00:09:58.360 | They can be used to manipulate other people's behavior.
00:10:00.360 | They can be used to manipulate other people's behavior.
00:10:02.360 | They can be used to manipulate other people's behavior.
00:10:04.360 | They can be used to manipulate other people's behavior.
00:10:06.360 | They can be used to manipulate other people's behavior.
00:10:08.360 | They can be used to manipulate other people's behavior.
00:10:10.360 | They can be used to manipulate other people's behavior.
00:10:12.360 | They can be used to manipulate other people's behavior.
00:10:14.360 | They can be used to manipulate other people's behavior.
00:10:16.360 | They can be used to manipulate other people's behavior.
00:10:18.360 | They can be used to manipulate other people's behavior.
00:10:20.360 | They can be used to manipulate other people's behavior.
00:10:22.360 | They can be used to manipulate other people's behavior.
00:10:24.320 | They can be used to manipulate other people's behavior.
00:10:26.320 | They can be used to manipulate other people's behavior.
00:10:28.320 | They can be used to manipulate other people's behavior.
00:10:30.320 | They can be used to manipulate other people's behavior.
00:10:32.320 | They can be used to manipulate other people's behavior.
00:10:34.320 | They can be used to manipulate other people's behavior.
00:10:36.320 | They can be used to manipulate other people's behavior.
00:10:38.320 | They can be used to manipulate other people's behavior.
00:10:40.320 | They can be used to manipulate other people's behavior.
00:10:42.320 | They can be used to manipulate other people's behavior.
00:10:44.320 | They can be used to manipulate other people's behavior.
00:10:46.320 | They can be used to manipulate other people's behavior.
00:10:48.320 | They can be used to manipulate other people's behavior.
00:10:50.320 | They can be used to manipulate other people's behavior.
00:10:52.320 | They can be used to manipulate other people's behavior.
00:10:54.280 | They can be used to manipulate other people's behavior.
00:10:56.280 | They can be used to manipulate other people's behavior.
00:10:58.280 | They can be used to manipulate other people's behavior.
00:11:00.280 | They can be used to manipulate other people's behavior.
00:11:02.280 | They can be used to manipulate other people's behavior.
00:11:04.280 | They can be used to manipulate other people's behavior.
00:11:06.280 | They can be used to manipulate other people's behavior.
00:11:08.280 | They can be used to manipulate other people's behavior.
00:11:10.280 | They can be used to manipulate other people's behavior.
00:11:12.280 | They can be used to manipulate other people's behavior.
00:11:14.280 | They can be used to manipulate other people's behavior.
00:11:16.280 | They can be used to manipulate other people's behavior.
00:11:18.280 | They can be used to manipulate other people's behavior.
00:11:20.280 | They can be used to manipulate other people's behavior.
00:11:22.280 | They can be used to manipulate other people's behavior.
00:11:24.240 | They can be used to manipulate other people's behavior.
00:11:26.240 | They can be used to manipulate other people's behavior.
00:11:28.240 | They can be used to manipulate other people's behavior.
00:11:30.240 | They can be used to manipulate other people's behavior.
00:11:32.240 | They can be used to manipulate other people's behavior.
00:11:34.240 | They can be used to manipulate other people's behavior.
00:11:36.240 | They can be used to manipulate other people's behavior.
00:11:38.240 | They can be used to manipulate other people's behavior.
00:11:40.240 | They can be used to manipulate other people's behavior.
00:11:42.240 | They can be used to manipulate other people's behavior.
00:11:44.240 | They can be used to manipulate other people's behavior.
00:11:46.240 | They can be used to manipulate other people's behavior.
00:11:48.240 | They can be used to manipulate other people's behavior.
00:11:50.240 | They can be used to manipulate other people's behavior.
00:11:52.240 | They can be used to manipulate other people's behavior.
00:11:54.200 | They can be used to manipulate other people's behavior.
00:11:56.200 | They can be used to manipulate other people's behavior.
00:11:58.200 | They can be used to manipulate other people's behavior.
00:12:00.200 | They can be used to manipulate other people's behavior.
00:12:02.200 | They can be used to manipulate other people's behavior.
00:12:04.200 | They can be used to manipulate other people's behavior.
00:12:06.200 | They can be used to manipulate other people's behavior.
00:12:08.200 | They can be used to manipulate other people's behavior.
00:12:10.200 | They can be used to manipulate other people's behavior.
00:12:12.200 | They can be used to manipulate other people's behavior.
00:12:14.200 | They can be used to manipulate other people's behavior.
00:12:16.200 | They can be used to manipulate other people's behavior.
00:12:18.200 | They can be used to manipulate other people's behavior.
00:12:20.200 | They can be used to manipulate other people's behavior.
00:12:22.200 | They can be used to manipulate other people's behavior.
00:12:24.160 | They can be used to manipulate other people's behavior.
00:12:26.160 | They can be used to manipulate other people's behavior.
00:12:28.160 | They can be used to manipulate other people's behavior.
00:12:30.160 | They can be used to manipulate other people's behavior.
00:12:32.160 | They can be used to manipulate other people's behavior.
00:12:34.160 | They can be used to manipulate other people's behavior.
00:12:36.160 | They can be used to manipulate other people's behavior.
00:12:38.160 | They can be used to manipulate other people's behavior.
00:12:40.160 | They can be used to manipulate other people's behavior.
00:12:42.160 | They can be used to manipulate other people's behavior.
00:12:44.160 | They can be used to manipulate other people's behavior.
00:12:46.160 | They can be used to manipulate other people's behavior.
00:12:48.160 | They can be used to manipulate other people's behavior.
00:12:50.160 | They can be used to manipulate other people's behavior.
00:12:52.160 | They can be used to manipulate other people's behavior.
00:12:54.120 | They can be used to manipulate other people's behavior.
00:12:56.120 | They can be used to manipulate other people's behavior.
00:12:58.120 | They can be used to manipulate other people's behavior.
00:13:00.120 | They can be used to manipulate other people's behavior.
00:13:02.120 | They can be used to manipulate other people's behavior.
00:13:04.120 | They can be used to manipulate other people's behavior.
00:13:06.120 | They can be used to manipulate other people's behavior.
00:13:08.120 | They can be used to manipulate other people's behavior.
00:13:10.120 | They can be used to manipulate other people's behavior.
00:13:12.120 | They can be used to manipulate other people's behavior.
00:13:14.120 | They can be used to manipulate other people's behavior.
00:13:16.120 | They can be used to manipulate other people's behavior.
00:13:18.120 | They can be used to manipulate other people's behavior.
00:13:20.120 | They can be used to manipulate other people's behavior.
00:13:22.120 | They can be used to manipulate other people's behavior.
00:13:24.080 | They can be used to manipulate other people's behavior.
00:13:26.080 | They can be used to manipulate other people's behavior.
00:13:28.080 | They can be used to manipulate other people's behavior.
00:13:30.080 | They can be used to manipulate other people's behavior.
00:13:32.080 | They can be used to manipulate other people's behavior.
00:13:34.080 | They can be used to manipulate other people's behavior.
00:13:36.080 | They can be used to manipulate other people's behavior.
00:13:38.080 | They can be used to manipulate other people's behavior.
00:13:40.080 | They can be used to manipulate other people's behavior.
00:13:42.080 | They can be used to manipulate other people's behavior.
00:13:44.080 | They can be used to manipulate other people's behavior.
00:13:46.080 | They can be used to manipulate other people's behavior.
00:13:48.080 | They can be used to manipulate other people's behavior.
00:13:50.080 | They can be used to manipulate other people's behavior.
00:13:52.080 | They can be used to manipulate other people's behavior.
00:13:54.040 | They can be used to manipulate other people's behavior.
00:13:56.040 | They can be used to manipulate other people's behavior.
00:13:58.040 | They can be used to manipulate other people's behavior.
00:14:00.040 | They can be used to manipulate other people's behavior.
00:14:02.040 | They can be used to manipulate other people's behavior.
00:14:04.040 | They can be used to manipulate other people's behavior.
00:14:06.040 | They can be used to manipulate other people's behavior.
00:14:08.040 | They can be used to manipulate other people's behavior.
00:14:10.040 | They can be used to manipulate other people's behavior.
00:14:12.040 | They can be used to manipulate other people's behavior.
00:14:14.040 | They can be used to manipulate other people's behavior.
00:14:16.040 | They can be used to manipulate other people's behavior.
00:14:18.040 | They can be used to manipulate other people's behavior.
00:14:20.040 | They can be used to manipulate other people's behavior.
00:14:22.040 | They can be used to manipulate other people's behavior.
00:14:24.000 | They can be used to manipulate other people's behavior.
00:14:26.000 | They can be used to manipulate other people's behavior.
00:14:28.000 | They can be used to manipulate other people's behavior.
00:14:30.000 | They can be used to manipulate other people's behavior.
00:14:32.000 | They can be used to manipulate other people's behavior.
00:14:34.000 | They can be used to manipulate other people's behavior.
00:14:36.000 | They can be used to manipulate other people's behavior.
00:14:38.000 | They can be used to manipulate other people's behavior.
00:14:40.000 | They can be used to manipulate other people's behavior.
00:14:42.000 | They can be used to manipulate other people's behavior.
00:14:44.000 | They can be used to manipulate other people's behavior.
00:14:46.000 | They can be used to manipulate other people's behavior.
00:14:48.000 | They can be used to manipulate other people's behavior.
00:14:50.000 | They can be used to manipulate other people's behavior.
00:14:52.000 | They can be used to manipulate other people's behavior.
00:14:53.960 | They can be used to manipulate other people's behavior.
00:14:55.960 | They can be used to manipulate other people's behavior.
00:14:57.960 | They can be used to manipulate other people's behavior.
00:14:59.960 | They can be used to manipulate other people's behavior.
00:15:01.960 | They can be used to manipulate other people's behavior.
00:15:03.960 | They can be used to manipulate other people's behavior.
00:15:05.960 | They can be used to manipulate other people's behavior.
00:15:07.960 | They can be used to manipulate other people's behavior.
00:15:09.960 | They can be used to manipulate other people's behavior.
00:15:11.960 | They can be used to manipulate other people's behavior.
00:15:13.960 | They can be used to manipulate other people's behavior.
00:15:15.960 | They can be used to manipulate other people's behavior.
00:15:17.960 | They can be used to manipulate other people's behavior.
00:15:19.960 | They can be used to manipulate other people's behavior.
00:15:21.960 | They can be used to manipulate other people's behavior.
00:15:23.920 | They can be used to manipulate other people's behavior.
00:15:25.920 | They can be used to manipulate other people's behavior.
00:15:27.920 | They can be used to manipulate other people's behavior.
00:15:29.920 | They can be used to manipulate other people's behavior.
00:15:31.920 | They can be used to manipulate other people's behavior.
00:15:33.920 | They can be used to manipulate other people's behavior.
00:15:35.920 | They can be used to manipulate other people's behavior.
00:15:37.920 | They can be used to manipulate other people's behavior.
00:15:39.920 | They can be used to manipulate other people's behavior.
00:15:41.920 | They can be used to manipulate other people's behavior.
00:15:43.920 | They can be used to manipulate other people's behavior.
00:15:45.920 | They can be used to manipulate other people's behavior.
00:15:47.920 | They can be used to manipulate other people's behavior.
00:15:49.920 | They can be used to manipulate other people's behavior.
00:15:51.920 | They can be used to manipulate other people's behavior.
00:15:53.880 | They can be used to manipulate other people's behavior.
00:15:55.880 | They can be used to manipulate other people's behavior.
00:15:57.880 | They can be used to manipulate other people's behavior.
00:15:59.880 | They can be used to manipulate other people's behavior.
00:16:01.880 | They can be used to manipulate other people's behavior.
00:16:03.880 | They can be used to manipulate other people's behavior.
00:16:05.880 | They can be used to manipulate other people's behavior.
00:16:07.880 | They can be used to manipulate other people's behavior.
00:16:09.880 | They can be used to manipulate other people's behavior.
00:16:11.880 | They can be used to manipulate other people's behavior.
00:16:13.880 | They can be used to manipulate other people's behavior.
00:16:15.880 | They can be used to manipulate other people's behavior.
00:16:17.880 | They can be used to manipulate other people's behavior.
00:16:19.880 | They can be used to manipulate other people's behavior.
00:16:21.880 | They can be used to manipulate other people's behavior.
00:16:23.840 | They can be used to manipulate other people's behavior.
00:16:25.840 | They can be used to manipulate other people's behavior.
00:16:27.840 | They can be used to manipulate other people's behavior.
00:16:29.840 | They can be used to manipulate other people's behavior.
00:16:31.840 | They can be used to manipulate other people's behavior.
00:16:33.840 | They can be used to manipulate other people's behavior.
00:16:35.840 | They can be used to manipulate other people's behavior.
00:16:37.840 | They can be used to manipulate other people's behavior.
00:16:39.840 | They can be used to manipulate other people's behavior.
00:16:41.840 | They can be used to manipulate other people's behavior.
00:16:43.840 | They can be used to manipulate other people's behavior.
00:16:45.840 | They can be used to manipulate other people's behavior.
00:16:47.840 | They can be used to manipulate other people's behavior.
00:16:49.840 | They can be used to manipulate other people's behavior.
00:16:51.840 | They can be used to manipulate other people's behavior.
00:16:53.800 | They can be used to manipulate other people's behavior.
00:16:55.800 | They can be used to manipulate other people's behavior.
00:16:57.800 | They can be used to manipulate other people's behavior.
00:16:59.800 | They can be used to manipulate other people's behavior.
00:17:01.800 | They can be used to manipulate other people's behavior.
00:17:03.800 | They can be used to manipulate other people's behavior.