The end of AI hype cycle?
My venture in AI-driven data analysis, thoughts on OpenAI developer day, and the future of AI as we know it.
You may have heard that OpenAI has just released a new set of features. While the internet is raving with joy, personally I think we are hitting a big slowdown. Why?
Think of OpenAI as the first iPhone. It was new, it was cool, and it was groundbreaking. Apple at the time was not the first company to build a smart phone, but they were the ones to get it just right.
Fast-forward to today and iPhone 15 Pro Max is still cool. It’s got lots of features, takes epic photos, and has a long battery life ..etc. It’s just not awesome anymore. It’s just an iPhone. There are in fact better Android phones out there now, but if you are on Apple, you are stuck in the iPhoto/iMessage ecosystem, and it is convenient, and there is no sense to switch.
The same is happening to Open AI. They have a fantastic product, and it will keep getting faster and smarted, but it does not seem like the next ChatGPT is coming any time soon. They have entered an optimization cycle.
X.ai , Google, Apple, and a handful of startups and open source projects are going to catch up, and in some cases surpass OpenAI very soon. New companies will emerge that will dive into niches, and they will shine and surpass generic AIs. A whole generation of companies will emerge just to connect all of this together and to help enterprises to make sense of all the available options. Life in AI is going to be great, but it is going to be a while until the next break through, and the next hype cycle. Let’s just hope OpenAI doesn’t go out the way of Blackberry.
How I am using AI to process large amounts of data.
I spent last week gathering more data and learning more Clojure.
What am I up to and why in the world am I learning a new programming language while ChatGPT can write it all for us?
Well, my friend Rafal is a “Sr. Clojure Developer” and I really like working with him, so in order to continue working together I had to compromise, and learn. But Clojure is actually a fun language, and it’s been very enjoyable for me. It’s fast, and great at data manipulation (will write about this next newsletter). The only challenge with learning Clojure was getting used to thinking about programming in a completely new way. Basically, my brain hurt until what used to be painful became beautiful. I wouldn’t recommend trying it, but I enjoyed it.
So what am I doing with it?
I am looking at way of reinventing Geekwire 200. It’s a list of 200 Seattle based companies, ranked using social signals.
It occurred to me that it would be very interesting to create this list of companies using not just any social data, but only Geekwire articles.
Was there enough content in news articles about these companies to pull out sentiment not just about how the public perceives them, but also about their finance, operations, and human resources, among other factors? How true is the information about these companies, if it’s only based on the news reports, to which those companies often contribute?
As it turned out after scanning over 2000 articles that yes, it was possible to turn that information into a company scorecard (see below), and as you can see, it’s quite different than the current version on the Geekwire site.
Now, is this data “good”?
That all depends on what you think defines good. Some people would say that if you take garbage in, you get garbage out, and I would agree. We don’t exactly know why Crowd Cow is a talent magnet, while Outreach likelihood of success is only slightly above average, while Common Room is leading as an overall winner. The way this model works is a black box, and we can ask it for an explanation, but we currently don’t have a way to verify that all the scores abide to the same criteria. The easiest way to understand why this is the case is to read “The Alignment Problem.”
Regardless of where the data comes from though, an how it is analyzed, as long as the data is consistently coming from the same place, and the model does its best to be consistent in evaluating each company, the signal will accurately respond to the changes.
Think about it like this - you can watch a car driving down a hill, and when the car accelerates, you can clearly see it. You don’t know if they are going faster because the driver is pushing the throttle, or because they’ve lost control of their breaks, but you know they are going faster.
In other words, given enough data, I think it is possible to predict the vector of startup success, whether or not the startup itself is spewing good news or not.
This, of course, is not a trivial problem at all. Right now I am only trolling through thousands of articles on Geekwire and this is already taking a rather long time. To do this better I’d need to scrape millions of data points across all of the web, and then process them in a meaningful way, ensure consistent results, and find a way to fine tune each metric to an acceptable standard.
If you think I should, or should not do more of this, let me know. Also, if you know of a company that can use data analysis like this, and wants to pay for it, I’d be happy to discuss.
The future of OpenAI and AIs in general.
Once again, last nigh internet erupted in discussions on what is possible with all the new features of OpenAI, primarily their app-store like GPTs, developer created instances of AI that can solve different niche problems. From their website:
Anyone can easily build their own GPT—no coding is required. You can make them for yourself, just for your company’s internal use, or for everyone. Creating one is as easy as starting a conversation, giving it instructions and extra knowledge, and picking what it can do, like searching the web, making images or analyzing data.Haven’t we seen it all before? iPhone App Store came around, enabled developers to create some money in the short run, but eventually apple just absorbed all the good ideas into their own apps, and now the only people making a killing on the store are big agencies making games. Most big companies have apps on the store to help interact with their products, but while helpful, those apps are not essential and can in most cases be replaced with a website.
Very few people in the valley invent a truly new game, most pick up the playbook and follow it faster and better in a new market.
That is why Sama is trying to scare everyone with the “killer AI,” so that your perception of what AI (and thus OpenAI) can do remains much higher than what it actually is. That is why OpenAI is essentially releasing mix tapes of the same songs; they got to keep the momentum going until they can figure out the next act.
That’s admirable, that’s awesome honestly, and that’s really hard, so all the kudos to them and I wish it works. I love using OpenAI and I want it to get better. But I also want it to get cheaper, and I want many more options without Microsoft control. I also want to run my own micro models that do anything I desire.
While the cost of training even a small model can run anywhere between tens of thousands of dollars to millions, my hope/bet is that personalized (and hopefully open source) models are going to be the way of the future.
I think that’s why Elon is creating a competitor to OpenAI. Partially because he can, partially because he’s pissed at Sama, partially because it’s a good business. Regardless, Xai (aka Grok) is going to become a big player.
https://twitter.com/elonmusk/status/1721795847136985337
Meanwhile, people are wondering - where are Apple and Google, how come they have not crushed OpenAI yet? The answer is they don’t have to. Those big guys have enough money to take their time. They are no longer in the business of being first, they are in the business of being last and long-lasing. At their size they don’t need to rush, as long as they can catch up. This is why OpenAI is running so fast, and this is why Satya Nadella showed up at the OpenAI developer conference yesterday. He was there to remind everyone who the real handler was.
AI as we know it is a great feature and I think every company that can find a way to integrate AI into their workflow should do so. It’s a multi-tool. However, as far as new companies being built around AI, I think those have to be either non technical companies that use AI to 10x their business, or they have to be deeply technical companies solving complicated problems. Everything in between will be quickly destroyed by the big companies, or diluted in value by the thousand others companies like them.
That’s it for today.
Got questions? Want to discuss, or maybe even build something together?
You can find my email here: kirillzubovsky.com
p.s. I leave you with this:
