Followups: BingGPT told Google to shut up and dance with me
The following highlights provided by Annotote: Don’t waste time and attention, get straight to the point. All signal. No noise.
The Roaring 20s: A future-proof economy and the business of adapting to automation/AI/ML
by Anthony Bardaro (Adventures in Consumer Technology) 2020.01.06
Microsoft researchers unveil Kosmos-1, a multimodal LLM they claim can understand image content, pass visual IQ tests, and accepts a variety of input formats
by Ars Technica 2023.03.01
OpenAI launches a ChatGPT API for businesses, with dedicated capacity plans, priced at $0.002/~750 words, and says Snap and Shopify are among the early adopters
by Tech Crunch 2023.03.01
Google’s search engine marketshare actually increased after Bing’s chat-based search launch
by Bernstein (hattip @carlquintanilla) 2023.03.01
despite $goog’s stock trading down -17% (>$225B of market cap lost) over the past three weeks in the wake of $msft’s ‘assault’ on their search business... Google GAINED Search Engine market share in February [2023.]
Google launches Workspace (formerly GSuite) with full AI/ML/LLM integrations–and the results look wow
by Google/YouTube (hattip @anthpb) 2023.03.15
Carthago delenda est
Related:
Google raises Workspace monthly pricing by 20% for new users
by 9to5Google 2023.03.14
See also:
Microsoft announces OpenAI (chatGPT/GPT-4) integrations for all of its productivity suite applications
by Microsoft (via YouTube) 2023.03.16
See also:
Adobe announces Firefly, which integrates generative AI LLMs to assist creators in the creation of images, videos, and text effects
by Adobe (via YouTube) 2023.03.22
The US Copyright Office says protections for AI-assisted artistic works depend on whether AI’s contributions reflect the author’s “own mental conception”
by Reuters 2023.03.16
The Copyright Office weighed in for the first time last month on whether its output is copyrightable, finding Midjourney-generated images in [a comic book] could not be protected, though [the author’s] text and unique arrangement of the book’s elements could.
The office reiterated Wednesday that copyright protection depends on the amount of human creativity involved, and that the most popular AI systems likely do not create copyrightable work.
“Based on the Office’s understanding of the generative AI technologies currently available, users do not exercise ultimate creative control over how such systems interpret prompts and generate material,” the office said. “Instead, these prompts function more like instructions to a commissioned artist”…
[I]ts policy “does not mean that technological tools cannot be part of the creative process… In each case, what matters is the extent to which the human had creative control over the work’s expression and actually formed the traditional elements of authorship,” the office said. The office also said that copyright applicants must disclose when their work includes AI-created material, and that previously filed applications that do not disclose AI’s role must be corrected.
Facing hardware shortages, Microsoft has had to ration GPU access for LLM developers since late 2022
by The Information 2023.03.16
Microsoft is facing an internal shortage of the server hardware needed to run the AI… That has forced the company to ration access to the hardware for some internal teams building other AI tools to ensure it has enough capacity to handle both Bing’s new GPT-4 powered chatbot and the upcoming new Office tools, set to be announced on Thursday. And the shortage of hardware may be affecting Microsoft customers[.]
On the question of sentience and Artificial General Intelligence: Continuous vs continual computing
by Anthony Bardaro (@anthpb via Twitter) 2023.03.27
Regarding probability and inference[:]
Narrow AI/LLMs right now are discrete — continual computing [compared to] the more continuous computing exhibited by organic general intelligence like the human brain…
[That’s inspired by] Jeff Hawkins’ “Theory of Intelligence”:
“The brain creates a predictive model. This just means that the brain continuously predicts what its inputs will be. Prediction isn’t something that the brain does every now and then; it is an intrinsic property that never stops, and it serves an essential role in learning. When the brain’s predictions are verified, that means the brain’s model of the world is accurate. A mis-prediction causes you to attend to the error and update the model.”
ChatGPT plugins are the answer to the LLM’s challenges with correctly and accurate answering certain questions (like math and science) that have objective answers
by Ben Thompson (Stratechery) 2022.03.27
Stephen Wolfram: “[T]he results are essentially never ‘perfect’. Maybe something works well 95% of the time. But try as one might, the other 5% remains elusive. For some purposes one might consider this a failure. But the key point is that there are often all sorts of important use cases for which 95% is ‘good enough’. Maybe it’s because the output is something where there isn’t really a ‘right answer’ anyway. Maybe it’s because one’s just trying to surface possibilities that a human — or a systematic algorithm — will then pick from or refine… And yes, there’ll be plenty of cases where ‘raw ChatGPT’ can help with people’s writing, make suggestions, or generate text that’s useful for various kinds of documents or interactions. But when it comes to setting up things that have to be perfect, machine learning just isn’t the way to do it — much as humans aren’t either… ChatGPT does great at the ‘human-like parts’, where there isn’t a precise ‘right answer’. But when it’s ‘put on the spot’ for something precise, it often falls down. But the whole point here is that there’s a great way to solve this problem — by connecting ChatGPT to Wolfram|Alpha and all its computational knowledge ‘superpowers’.”
That’s exactly what OpenAI has done. From The Verge:
“OpenAI is adding support for plug-ins to ChatGPT — an upgrade that massively expands the chatbot’s capabilities and gives it access for the first time to live data from the web. Up until now, ChatGPT has been limited by the fact it can only pull information from its training data, which ends in 2021. OpenAI says plug-ins will not only allow the bot to browse the web but also interact with specific websites, potentially turning the system into a wide-ranging interface for all sorts of services and sites. In an announcement post, the company says it’s almost like letting other services be ChatGPT’s ‘eyes and ears’.”
Stephen Wolfram’s Wolfram|Alpha is one of the official plugins, and now ChatGPT gets the above answers right — and quickly [including a code-based audit trail of its API calls back-and-forth with the plugin service]… Right now there are 11 plugins in categories like Travel (Expedia and Kayak), restaurant reservations (OpenTable), and Zapier, which opens the door to 5,000+ other apps (the plugin to search the web isn’t currently available); they are all presented in what is being called the “Plugin store”.
Meta CTO Andrew Bosworth says the company plans to begin commercializing its generative AI (LLaMa) in 2023 and remains at the “very forefront” of the LLM field
by Nikkei Asia 2023.04.05
Bosworth believes Meta’s artificial intelligence can improve an ad’s effectiveness partly by telling the advertiser what tools to use in making it. He said that instead of a company using a single image in an advertising campaign, it can “ask the AI, ‘Make images for my company that work for different audiences.’ And it can save a lot of time and money”…
The technology will also be used in the metaverse... “So previously,” Bosworth said, “if I wanted to create a 3D world, I needed to learn a lot of computer graphics and programming. In the future, you might be able to just describe the world you want to create and have the large language model generate that world for you. And so it makes things like content creation much more accessible to more people.”
Google announces data on its 4th-generation TPU-based supercomputers, used for AI training
by Reuters 2023.04.04
Google on Tuesday published a scientific paper detailing how it has strung more than 4,000 of the [Tensor Processing Unit] chips together into a supercomputer using its own custom-developed optical switches to help connect individual machines…
Google claims its 4th-generation TPU-based supercomputers, used for AI training, are up to 1.7x faster and 1.9x more power-efficient than Nvidia’s A100 systems…
Google said it did not compare its fourth-generation to Nvidia’s current flagship H100 chip because the H100 came to the market after Google’s chip and is made with newer technology [but] Google hinted that it might be working on a new TPU that would compete with the Nvidia H100.
Leaked internal memo from senior Google engineer: “We have no moat, and neither does OpenAI”
by anon (via Dylan Patel/Semianalysis) 2023.05.04
[T]he uncomfortable truth is, we aren’t positioned to win this arms race and neither is OpenAI. While we’ve been squabbling, a third faction has been quietly eating our lunch. I’m talking, of course, about open source. Plainly put, they are lapping us. Things we consider “major open problems” are solved and in people’s hands today. Just to name a few:
• LLMs on a Phone: People are running foundation models on a Pixel 6 at 5 tokens [per second].
• Scalable Personal AI: You can finetune a personalized AI on your laptop in an evening.
• Responsible Release: This one isn’t “solved” so much as “obviated”. There are entire websites full of art models with no restrictions whatsoever, and text is not far behind.
• Multimodality: The current multimodal ScienceQA SOTA was trained in an hour.
While our models still hold a slight edge in terms of quality, the gap is closing astonishingly quickly. Open-source models are faster, more customizable, more private, and pound-for-pound more capable. They are doing things with $100 and 13B params that we struggle with at $10M and 540B. And they are doing so in weeks, not months. This has profound implications for us:
• We have no secret sauce. Our best hope is to learn from and collaborate with what others are doing outside Google. We should prioritize enabling 3P integrations.
• People will not pay for a restricted model when free, unrestricted alternatives are comparable in quality. We should consider where our value add really is.
• Giant models are slowing us down. In the long run, the best models are the ones which can be iterated upon quickly. We should make small variants more than an afterthought, now that we know what is possible in the <20B parameter regime.
LoRA is an incredibly powerful technique we should probably be paying more attention to… [It] reduces the size of the update matrices by a factor of up to several thousand. This allows model fine-tuning at a fraction of the cost and time. Being able to personalize a language model in a few hours on consumer hardware is a big deal… in terms of engineer-hours, the pace of improvement from these models vastly outstrips what we can do with our largest variants, and the best are already largely indistinguishable from ChatGPT… Who would pay for a Google product with usage restrictions if there is a free, high quality alternative without them? […] The legal cover afforded by “personal use” and the impracticality of prosecuting individuals means that individuals are getting access to these technologies while they are hot…
Much of this innovation is happening on top of the leaked model weights from Meta [LLaMA]… the one clear winner in all of this is Meta [Facebook]. Because the leaked model was theirs, they have effectively garnered an entire planet’s worth of free labor. Since most open source innovation is happening on top of their architecture, there is nothing stopping them from directly incorporating it into their products.
Google I/O 2023 developer conference: The coming-out party
by Ben Thompson (Stratechery) 2023.05.15
From [CEO Sundar] Pichai’s opening statement:
“Seven years into our journey as an AI-first company, we’re at an exciting inflection point… We’ve been applying AI to make our products radically more helpful for a while. With generative AI, we’re taking the next step. With a bold and responsible approach, we are re-imagining all of our core products, including search.” […]
Google is in it to win it [and] taking a hybrid approach: most searches are not commercial, and so Google is going to place generated text right at the top… For searches that do have commercial possibilities, ads will still get top billing… This seems like a reasonable approach, aided by the fact that non-commercial searches are probably more likely to benefit from AI anyways; this is also an approach that already looks more compelling and yes, bold, than Microsoft’s grafting on of Bing Chat to the side of traditional search…
I take this Google I/O as evidence that AI is in fact a sustaining technology for all of Big Tech, including Google. Moreover, if that is the case, then that is a reason to be less bearish on the search company, because all of the reasons to expect them to have a leadership position — from capabilities to data to infrastructure to a plethora of consumer touch points — remain. Still, the challenges facing search as presently constructed — particularly its ad model — remain.
Google plans to use generative AI to boost the relevance of Search ads based on query context and debuts a conversational experience for creating ad campaigns
by TechCrunch 2023.05.23
“Simply add a preferred landing page from your website and Google AI will summarize the page,” Google explained in a blog post. “Then, it will generate relevant and effective keywords, headlines, descriptions, images and other assets for your campaign. You can review and easily edit these suggestions before deploying. Now, you can chat your way into better performance — ask Google AI for ideas, just like you might ask a colleague.”
Google [also] announced that it’s launching Product Studio, a new tool that lets merchants easily create product imagery using generative AI… gives merchants the ability to create product imagery for free and get more value from the images they already have. You can create new imagery without the added cost of new photoshoots.
See also:
Google plans to experiment with running Search and Shopping ads, which are fully customized based on users’ queries, inside its conversational AI Search tool
by TechCrunch 2023.05.23
ChatGPT, generative AI, and how/when the automation of work will (and won’t) change the workplace
by Benedict Evans 2023.07.02
If you do understand [that LLMs are probabilistic pattern matching and not deterministic authoritative answering], then you have to ask, well, where are LLMs useful? Where is it useful to have automated undergraduates, or automated interns, who can repeat a pattern, that you might have to check? The last wave of machine learning gave you infinite interns who could read anything for you, but you had to check, and now we have infinite interns that can write anything for you, but you have to check. So where is it useful to have infinite interns?
[W]e’re back to the Jevons Paradox… we’ve been automating work for 200 years. Every time we go through a wave of automation, whole classes of jobs go away, but new classes of jobs get created. There is frictional pain and dislocation in that process, and sometimes the new jobs go to different people in different places, but over time the total number of jobs doesn’t go down, and we have all become more prosperous…
The Lump of Labour fallacy is the misconception that there is a fixed amount of work to be done, and that if some work is taken by a machine then there will be less work for people. But if it becomes cheaper to use a machine to make, say, a pair of shoes, then the shoes are cheaper, more people can buy shoes and they have more money to spend on other things besides, and we discover new things we need or want, and new jobs. The efficient gain isn’t confined to the shoe: generally, it ripples outward through the economy and creates new prosperity and new jobs…
Meanwhile, while one could suggest that LLMs will subsume many apps on one axis, I think it’s equally likely that they will enable a whole new wave of unbundling on other axes, as startups peel off dozens more use cases from Word, Salesforce and SAP, and build a whole bunch more big companies by solving problems that no-one had realised were problems until LLMs let you solve them. [The] typical customer now has hundreds of different software applications, and enterprise customers have almost 500. And yet, enterprise cloud adoption is still barely a quarter of workflows. What does that mean for generative AI in the workplace? Whatever you think will happen, it will take years, not weeks…
This takes me, obviously, to AGI… [W]hat would happen if we had a system that didn’t have an error rate, didn’t hallucinate, and really could do anything that people can do[?] If we had that, then you might not have one accountant using Excel to get the output of ten accountants: you might just have the machine. This time, it really would be be different. Where previous waves of automation meant one person could do more, now you don’t need the person. Like a lot of AGI questions, though, this can become circular if you’re not careful. ‘If we had a machine that could do anything people do, without any of these limitations, could it do anything people can do, without these limitations?’ […] I tend to prefer Hume’s empiricism over Descartes — I can only analyse what we can know. We don’t have AGI, and without that, we have only another wave of automation, and we don’t seem to have any a priori reason why this must be more or less painful than all the others.
Interview: Alphabet’s lead in the AI race and AlphaFold as the first commercially viable, practical, useful application thereof (Demis Hassabis, Google AI/DeepMind, CEO)
by Nilay Patel (The Verge/Decoder podcast) 2023.07.10
[T]he large language models [LLM] have really entered the public consciousness because it’s something the average person… can actually understand and interact with. And, of course, language is core to human intelligence and our everyday lives. I think that does explain why chatbots specifically have gone viral in the way they have. Even though I would say things like AlphaFold [had] actually had the most unequivocally biggest beneficial effects so far in AI on the world because if you talk to any biologist or there’s a million biologists now, researchers and medical researchers, have used AlphaFold. I think that’s nearly every biologist in the world. Every Big Pharma company is using it to advance their drug discovery programs. I’ve had multiple, dozens, of Nobel Prize-winner-level biologists and chemists talk to me about how they’re using AlphaFold.
So a certain set of all the world’s scientists, let’s say, they all know AlphaFold, and it’s affected and massively accelerated their important research work. But of course, the average person in the street doesn’t know what proteins are even and doesn’t know what the importance of those things are for things like drug discovery…
I mean, AlphaFold was open sourced, right? So we obviously believe in open source and supporting research and open research. That’s a key thing of the scientific discourse, which we’ve been a huge part of. And so is Google, of course, publishing transformers and other things. And TensorFlow and you look at all the things we’ve done.
Previously:
Perplexity: An AI native answer engine to challenge Google Search
by The Wall Street Journal (WSJ) 2024.01.04
Perplexity, a startup going after Google’s dominant position in web search, has won backing from Jeff Bezos [who was a seed investor in Google] and venture capitalists [all of whom invested] $74M in the [round to value the company] at $520M…
[T]heir advantage is using advances in AI to provide direct answers, instead of website links, in response to search queries.. Perplexity maintains its own index of webpages, which it combines with a mixture of AI technology it has designed itself and purchased from outside providers such as OpenAI…
$5M to $10M in annual revenue from subscriptions and selling its AI software to other businesses… The company charges $20 a month for a more powerful version of the search engine that uses GPT-4, OpenAI, [and Perplexity] hasn’t ruled out introducing ads in the future…
Started less than two years ago, Perplexity has fewer than 40 employees and is [still] based out of a San Francisco co-working space. The company’s product, which it calls an answer engine, is used by about 10M people monthly… The startup has spent almost nothing on traditional marketing, relying on word of mouth and buzz on X to attract new users [yet] Perplexity’s website and mobile apps had 53M visits in November, up from 2.2M when the service became available in December 2022…
Neeva, a search startup that used generative AI to provide direct answers, shut down last year after it failed to gain enough traction[.]
Apple is negotiating a deal with Google to license Gemini AI engine for integration into iPhones and iOS 18, where they’d run on-device
by Anthony Bardaro (@anthpb via Twitter, hattip Bloomberg and The Verge) 2024.03.18
Apple is in talks to build Google’s Gemini AI engine into the iPhone in a potential blockbuster deal [and] considered OpenAI after proprietary model Apple GPT Ajax model underwhelmed
See also:
Apple Releases AI Research Paper (MM1) and Apple joining forces with Google via Gemini license?
by Ben Thompson (stratechery) 2024.03.18
Apple doesn’t appear to be anywhere close [to training their own MM1 model. Even] if Apple had their own Gemini-level model, they do not have the infrastructure to serve it to their massive user base, and if they wanted to build said infrastructure, they would need to get in line for Nvidia GPUs behind everyone else — and keep in mind that Apple and Nvidia have been giving each other the cold shoulder for over a decade, thanks in part to a never-settled dispute about Nvidia chip issues in Apple laptops. I doubt that Nvidia CEO Jensen Huang would be willing to do the company any favors, even if they wanted to build an inference cloud for serving their own models, which don’t exist.
Indeed, it’s possible that there isn’t an inference cloud that is large enough to service the iPhone user base outside of Google: ChatGPT can still occasionally struggle under high usage, and that’s for people who know to go and use it; being built into iOS has the potential to unlock an entirely new level of usage that only Google, with its TPU-based infrastructure spread all over the world, can even hope to handle (to the extent there were talks with OpenAI and Anthropic, I suspect they were about being available as an option).
Third, Gemini’s embarrassing launch is, from Apple’s perspective, a bit of a feature, not a bug. First off, it’s a reminder to the company of the risk entailed in having your own LLM: you own the output. If Gemini is doing the text and image generation, though, well, that’s Google’s fault! It’s kind of how Apple can benefit from the wide-scale user tracking done by Google — while condemning such wide-scale user tracking — by being paid by Google for the default position in search. Moreover, the Gemini mess may very well have made this laundry service cheaper to acquire…
[I]t’s not out of the question that Apple might pay for access (or negotiate a credit against the money it gets from Google for search results)…
The ultimate result, should this deal come to fruition, would probably be a hybrid: Apple will focus on smaller models that it can run on its devices, ideally with an API that makes certain functionality available to developers, who will no longer need to download a model independently. Those models will probably not be surfaced to users directly, but rather be built into products like Siri. More complicated queries, meanwhile, like text generation, will be sent to the cloud to be handled by Gemini. Indeed, this is how Siri works today: it handles basic queries and hands most off to search engines.
Several legal experts say Section 230 will not protect firms from lawsuits over the outputs of generative AI, echoing SCOTUS’s 2023 statement
by The Wall Street Journal (WSJ) 2024.03.30
I spoke with several legal experts across the ideological spectrum, and none expect that Section 230 will protect companies from lawsuits over the outputs of generative AI, which now include not just text but also images, music and video…
Among the most compelling indications that companies won’t be protected by current law is Supreme Court Justice Neil Gorsuch’s early 2023 statement on the subject… “Artificial intelligence generates poetry. It generates polemics today that would be content that goes beyond picking, choosing, analyzing or digesting content. And that is not protected.”
Google earnings call (2024q1)
by Mostly Borrow Ideas (MBI Deep Dives) 2024.04.25
[Sundar Pichai, Google CEO:] if you were to step back at this moment, there were a lot of questions last year, and we always felt confident and comfortable that we would be able to improve the user experience. People question whether these things would be costly to serve, and we are very, very confident we can manage the cost of how to serve these queries. People worried about latency. When I look at the progress we have made in latency and efficiency, we feel comfortable. There are questions about monetization. And based on our testing so far, I’m comfortable and confident that we’ll be able to manage the monetization transition here well as well. It will play out over time, but I feel we are well positioned. And more importantly, when I look at the innovation that’s ahead and the way the teams are working hard on it, I am very excited about the future ahead.
See also:
[Sundar Pichai, Google CEO:] we are very focused on our cost structures, procurement, and efficiency. And a number of technical breakthroughs are enhancing machine speed and efficiency, including the new family of Gemini models and a new generation of TPUs. For example, since introducing SGE about a year ago, machine costs associated with SGE responses have decreased 80% from when first introduced in labs driven by hardware, engineering, and technical breakthroughs. We remain committed to all of this work. Finally, our monetization path. We have clear paths to AI monetization through ads and cloud, as well as subscriptions. Philip will talk more about new AI features that are helping advertisers, including bringing Gemini models into Performance Max.
DOJ evidence: Microsoft invested in OpenAI over fears of falling behind Google
by Bloomberg 2024.04.30
Contrary to popular belief, Microsoft was playing catchup with Google, not the other way around:
Microsoft’s motivation for investing heavily and partnering with OpenAI came from a sense of falling badly behind Google, according to an internal email released Tuesday as part of the Justice Department’s antitrust case against the search giant. [MSFT’s CTO], Kevin Scott, was “very, very worried” when he looked at the AI model-training capability gap between Alphabet’s efforts and Microsoft’s, he wrote in a 2019 message to CEO Satya Nadella and co-founder Bill Gates… they lacked the infrastructure and development speed to catch up to the likes of OpenAI and Google’s DeepMind… “We are multiple years behind the competition in terms of machine learning scale,” Scott said in the email.
The first knowledge network
Annotote gives you highlights of everything you need to read and lets you annotate anything you want to save or share. Check out the most frictionless way to get informed or inform others:
