Google is losing control • TechCrunch

Google is upset. After years of single-minded worship of false gods, virtual assistants, the company is rushing its AI strategy as competitors join hands and raise the pitchfork. , it’s all happening because Google thought it had cornered the Pitchfork market.

In 2017, Google researchers published an article “Attention is all you need”, introducing the concept of transformers and significantly improving the capabilities of machine learning models. You don’t have to know the technical side (actually, I’m not in a position to teach). Suffice it to say that it is the T for GPT.

Why did Google offer this great stuff for free? Large private research institutions have been criticized in the past for withholding research, but the trend in recent years has been towards publication. This is a game of honor and a concession to the researchers themselves. There is probably an element of arrogance to it as well. How come you invented the technology and Google didn’t make the most of it?

The features we see in ChatGPT and other large language models today didn’t come immediately. It takes time to understand and take advantage of new tools. All major technology companies have had to consider what the new era of AI has to offer and what it requires.

assist the assistant

There is no doubt that Google, like any other company, was dedicated to AI work. Over the next few years, we made great strides in designing AI computational hardware, building a convenient platform for developers to test and develop machine learning models, from fine-tuning esoteric models to more advanced techniques such as speech synthesis. He published a ton of papers on everything, even the most recognizable.

However, there was a problem. I’ve heard this anecdotally from Google employees and industry insiders, but there’s a kind of feudal aspect to the company’s way of doing things. Having projects under the auspices of existing leading products such as Maps and Assistant is a reliable way to go. get money and personnel So, despite attracting many of the world’s best AI researchers, their talents seem to have been channeled into the ruts of corporate strategy.

Let’s see how it went? Here’s a (admittedly selective) little timeline:

In 2018, Google Assistant flows, photography (such as colorizing monochrome images), smart displays with a “visual-first version of the assistant” (have you seen it?), map assistants, AI-assisted, and more. showed significant improvement. Google News and (to their credit) MLKit.

2019 includes rebranded larger Smart Displays, AR Search Results, AR Maps, Google Lens updates, Duplex for the web (remember Duplex?), Compressed to do more locally Google Assistant, Waze Assistant, Driving Mode Assistant, Live Captioning and Live Relay (Speech Recognition) and projects to better understand people with speech impairments.

Certainly some of these are great! However, most were existing, but with AI boosts. In retrospect, many feel a little disgusted. It’s easy to see how big companies like Google act on and drive trends.

Meanwhile, in February of that year, I also ran the headline: “OpenAI built a very good text generator, but it’s considered too risky to release.” That was GPT-2. Not 3, not 3.5… 2.

In 2020, Google cloned AI-powered Pinterest and in December fired Timnit Gebru, one of the leading AI ethics leaders, over a paper pointing out the limitations and dangers of technology. .

To be fair, 2020 hasn’t been a great year for many — OpenAI’s notable exception is what co-founder Sam Altman had to do. Personally quell the GPT-3 hype Because it was more than I could bear.

In 2021, Google’s own large language model, LaMDA, debuted, but it didn’t sell well in demos. Perhaps you were still casting for the reason it exists other than making the assistant throw fewer errors.

OpenAI kicked off the year by showing off DALL-E, the first version of its soon-to-be-famous text-to-image model. Through systems like CLIP, they have begun to show that LLM can do more than language tasks. (To be clear, I don’t mean “artificial intelligence” or AGI, just that the process worked for more than a preconfigured collection of verbal commands.)

2022 will see more tweaks to Assistant, smart display enhancements, AR enhancements in Maps, and a $100 million acquisition of AI-generated profile pictures. OpenAI released his DALL-E 2 in April and ChatGPT in December.

I suspect that at some point in early 2022, Google executives woke up and were horrified by what they saw. I envision a scene in Lord of the Rings where Denethor finally looks over Mordor’s assembled army. But instead of losing their minds and being placed in wizardry, these desperate VPs sent emails asking why some startups are running circles around the world leader in AI. Especially after they’ve actually invented the means.

Proof of this is Imagen’s trot a month after DALL-E 2. However, like almost all other interesting AI research published by Google, no one was able to test it, let alone connect to the API. Then, after Meta released his Make-A-Video in September, Google responded with Imagen Video a week after him. Riffusion made waves in music production, and a month later came MusicLM (which is unusable).

But it was definitely ChatGPT that quickly moved Google’s management from trepidation to utter sweat flops.

It was clear to everyone involved that this kind of conversational AI would be very different from the assistant products Google has been investing in for 10 years. actual Run everyone else’s pseudo-AI (effectively a natural language front-end to a collection of APIs) pretended To. It’s a so-called life-and-death crisis.

Luck or foresight?

Now, the fact that someone from an acquisition-unaffected startup has sparked the next evolutionary phase of the search engine and has done it in a very public way that has captured the imagination of industry leaders and ordinary people alike. It was bad enough. technical avoidance.The real twist in the knife came from an unexpected place microsoft.

Calling Bing a “rival” to Google search is probably too generous. With about 3% of global searches compared to Google’s 92%, Bing is the wealthy gadfly. Microsoft seems to have abandoned its illusions about its ability to improve Bing’s standing and turned to help outside of its own home. Whether their investment in his OpenAI was a supernatural foresight or a happy accident, at some point it became clear they were backing a fast horse.

In a supposedly smoke-filled room, Satya Nadella and Sam Altman conspired to remove Google from the New World Order. Whatever the story behind it, Microsoft has secured loyalty with an innovative newcomer and an opportunity to leverage its technology where it can be most effective.

I’ve seen some interesting ideas pop up about how generative AI can help with productivity, coding, and even management, but I’ve seen copyright issues and the tendency of the AI ​​to react a little too ‘creatively’. has not yet been proven because of But with the right guardrails, it was clear that they were very good at synthesizing information to answer just about anything, from simple factual questions to complex philosophical questions.

Search combines the need for innovation to advance in the core competencies of language models at scale, which by chance or common sense just lined up the world’s leading authors as partners. The move to integrate the latest GPT model (some people call it GPT-4, but I suspect OpenAI has reserved that moniker for its own first-party model) into Bing and Edge is sort of It’s like a forced hail, and its last and best play. The world of search engines.

The day before Microsoft was slated for its big event to announce OpenAI-powered Bing, Google was visibly shaken and attempted a spoiler campaign with an empty blog post. Bard was apparently the name of his Google’s LaMDA-based ChatGPT competitor, now announced in a popular understated form. It promises features and has no hard dates or access plans.

This announcement attempt seemed so rushed that it was barely mentioned at Google’s Search and AI event two days later. It was promoting the future of the Knowledge Graph. The imagery used to describe Byrd contained a significant error that the James Webb Space Telescope “taken the first picture of a planet outside our solar system.” This is not true. The fact that this vaunted machine intelligence was wrong, and that no one at Google was paying enough attention to notice and check it, seems to have taken investors by surprise.

ChatGPT certainly has its problems, and shortly after Microsoft’s enhanced Bing went public, TechCrunch took what seemed safe and appropriate AI to improvise Hitler’s essays and tweak its own predecessors. was able to regurgitate the vaccine disinformation a version wrote last month. But these are the blemishes of an established record that includes billions of prompts and dialogue provided and overwhelming user satisfaction.

Google’s rush to backtrack on its attack clearly shows that it is unprepared, even at a limited experimental level, let alone a global rollout like the one Microsoft has already begun. It tells the story.

In a call to investors, CEO Sundar Pichai said: It’s early days, but you can see that we’re boldly announcing things, getting feedback, iterating, and improving things. ’ Does that sound like a man with a plan?

I can understand Google not wanting to slaughter the Golden Goose by prematurely merging half-baked generic LLMs and Search around Google. They have become experts in deploying highly specialized AI task models that do one or two things. However, their comfortable position has been plagued by inertia when it comes to making large movements.

Is it Google’s downfall? Of course not, it remains in default and for the foreseeable future is a ridiculously profitable and somewhat ridiculous company. Over the past few years, it turns out that Google’s failure to innovate meaningfully may have been due to reticence and pride, not wisdom or confidence. (It doesn’t help that the FTC and Justice are giving that advertising business another shot.)

However, this turn of the worm is only the first few, and we shouldn’t overestimate when the technology in question has yet to prove as valuable as everyone believes. Otherwise, the entire tech industry will be affected, not just Google.