AI Reading for Wednesday May 15! Google I/O
Google introduced and demoed a whole lot of stuff:
Project Astra. Google’s version of the assistant OpenAI demoed yesterday. Also a smart-glasses product is being demoed in the second part, worn in the inset and alluded to but not clearly called out. Gemini Live, a production version this summer, probably not the whole shebang.
Model news
Gemini Advanced, their $20/month LLM product, is now running on Gemini 1.5 Pro with 1m-token context window, which is like 1500 pages. This is the mixture-of-agents version of Gemini 1.0. It is essentially tied with GPT-4 and Anthropic Claude. In waitlist-gated preview is the expanded 2m-token context (3000 pages).
1.5 Flash, a smaller faster lower-latency lower-cost model that can presumably run on a single TPU.
Gemma, a 27b-parameter open-source model which is claimed to be comparable to Llama-3.
PaliGemma for images.
Nano Gemini integrated in Chrome for completion and whatnot.
LearnLM for education.
Gems, which sounds somewhat like OpenAI’s GPTs / GPT Store. Also demoed an experience like OpenAI’s Advanced Data Analysis.
Veo, Google’s answer to Sora, with a movie by Donald Glover. VideoFX is available now.
Imagen 3, Google’s answer to Midjourney and Dall-E
Music AI Sandbox (no writing lyrics and singing them like Sora, yet).
Trillium, a new TPM chip, which they say is 4.7x faster and 67% more power efficient and can run in 256-chip pods. I’d be curious how it competes with e.g. Nvidia B200 in TFLOPs, TFLOPs/w, RAM/how big a model you can run on a single chip.
Search stuff (wait does Google still do search or just AI?) , AI overviews, chat-type questions like make a meal plan, ask questions about a video of a meeting, AI photo search, AI Map search, trip planning, summarize/ask questions about your mail like summarize all the contractor bids.
Agent stuff. Help me return this item, help me manage receipts and whatnot. AI teammates in their version of Slack/Teams.
A Google IDE with AI.
NotebookLM, a note taker / educational product with AI to summarize, explain, chat with lecture notes presumably using the LearnLM fine-tuned model.
Takeaways.
Shock and awe…OpenAI awakened a sleeping giant.
The long context and multimodality of Gemini are underhyped. Ask it to summarize a whole book, or make an assistant that has 8 of your data science books, without RAG. It might run up a bit of a bill but it’s a gamechanger.
The Astra and OpenAI assistant demos might be a bit overhyped. We’ll see. I haven’t seen Gemini quite understand video as well as these demos, they might be cherry-picked.
Yesterday I talked about analogies to the dot-com era. One company that is no longer around is Netscape. They invented the Web as we know it. Microsoft really ground them down and pounded them into the ground.
Here we see what happens when Google gets a Halloween memo and really goes all-in. They are about organizing the world’s information. They are about machine learning. They did the first deep learning NLP products with Translate. They invented transformers. They had a model that a guy freaked out and thought was sentient, before anyone heard about OpenAI. They did AlphaGo and AlphaFold. This is a moment they have been training for their whole existence. And somehow they got caught napping by OpenAI and couldn’t ship anything that made them relevant for over a year. Well they are relevant now. OpenAI is now asking, ‘who are these guys?’
Google has control of a full ecosystem with billions of users, their own integrated cloud platform, their own mobile platform, and 60,000 engineers. Sir Demis and Jeff Dean are very credible AI leaders and Google can do stuff at scale, if not exactly nimbly. OpenAI has like a thousand employees total.
Netscape did some awesome stuff. But they also invented the blink tag, and JavaScript, the blink tag of programming languages. Once Microsoft came out with IE4 and bundled it with Windows, and Apache emerged as the web server standard, they were competing with superior products that were also free. Browsers and web servers turned out to be a feature of a platform ecosystem more than a product.
OpenAI is still a superior product, when you look at the model and the overall platform, if not by nearly as much as 6 months ago.
But it’s not 1999. Microsoft got their wings clipped by an antitrust suit. Google has to be mindful of that. And anyway Google doesn’t have a Microsoft type of monopoly now. OpenAI has Microsoft in their corner (even if they are hedging their bets with Suleyman). Open AI has a deal with Apple for mobile integration. Those are the best distribution partners OpenAI could want for enterprise and consumer respectively. But the setup could be unwieldy with a lot of cooks and OpenAI’s own governance issues and internal drama.
The ball is in OpenAI’s court. I want to see GPT-5 and Q* and long context pretty soon. The playing the raindrops era is over and if they don’t keep moving fast they might end up like Butch and Sundance.
In other news:
Ilya is out. - The Verge
No surprise but the risk of safety taking a back seat to competitive pressures is overwhelming. Hopefully no foom ‘fast onset of overwhelming mastery’.
Gemini 1.5 Flash watches itself being announced. - Twitter
Gemini also watched the OpenAI announcement, said it was impressed. The future is weird. - Twitter
What is in the library at OpenAI headquarters? - NYT
AI will hit the labor market like a ‘tsunami,’ says IMF chief. Reassuring. - Fortune
AWS CEO out after 3 years with mixed reviews. - Business Insider
Can you replace a Republican political consultant with AI? CNN asking the tough questions. - CNN
Testing LLMs on coding problems, we see a huge gap between success on well-known problems and novel problems. - HackerNoon
Conspicuously missing from Google’s announcements: AlphaCode 2 stuff. Last year they claimed it beat human competitive programmers. That would be a huge breakthrough. A lot of people didn’t believe that and thought the training data might contaminated. The Code Assist product they ship is not notably better than GitHub Copilot, which is pretty amazing. There are a lot of teaser demos that never become ready for prime time.
Everyone's making their own AI chip now, including a lot of folks not mentioned here. - The Verge
Estimates predict $70k for Nvidia B200 chips, a full server rack will run you $3m, not including the nuclear plant to power it. - Tom's Hardware
Gary Marcus has thoughts and spin. - Gary Marcus
In the short run the naysayers might score some points, but let’s stay focused on the long run, where we are truly utterly doomed.
Kevin Roose on a possible new ‘Her’ era of AI Samanthas - NYT
Llama-3: I invented Gemini! Yeah, that’s the ticket. - Reddit
An 'affordable' $16k humanoid robot - - YouTube
Follow the latest AI headlines via SkynetAndChill.com on Bluesky