Bing around and find out • TechCrunch

Microsoft’s new A revamped Bing with a custom version of OpenAI’s ChatGPT experienced a dizzyingly rapid reversal. And, well, it’s all Microsoft’s fault.

ChatGPT is a very interesting demonstration of new and unfamiliar technology. So, like all other AI adjacency structures to come, the novelty will cause everyone, from strong tech types to people who aren’t normally interested in the field, to overestimate its capabilities. That’s not surprising.

We are at a good “technical readiness level” to discuss over tea or beer. What are the benefits and risks of generative AI for art, literature, or philosophy? How can you be sure it’s original, imitation, hallucination? What are the implications for creators, coders, and customer service representatives?Finally, after his two years in crypto, let’s talk about something interesting.

The hype seems too big because it’s a technology designed to provoke more or less controversy, borrowing from the controversy common to all AI advancements. It’s similar to “The Dress” in that regard. command Responses, and those responses, generate further responses. Hype is, in a way, self-generated.

More than just arguments, large-scale language models like ChatGPT are also well-suited for low-stakes experiments such as Mario Never Ending. In fact, this is the fundamental approach to OpenAI development. We’ll release the model privately first to polish the sharpest edges, then publish it publicly to see how it reacts when a million users kick a tire at the same time. At some point people will give you money.

nothing to gain, nothing to lose

The key to this approach is that “failures” don’t actually have negative consequences, only positive consequences. By characterizing the model as experimental and even academic, participation or involvement in the GPT series of models is merely a large-scale test.

When someone builds something cool, it reinforces the idea that the model is promising. What else did someone actually expect from an experimental AI if they found a noticeable failure condition? It sinks into obscurity. If everything is so, then nothing unexpected — the miracle is that the model works just like that.

In this way, OpenAI has collected an astonishing amount of unique test data to refine its models. Millions of people have poked and poked his GPT-2, GPT-3, ChatGPT, DALL-E, and DALL-E 2 (among others) to find out their features, drawbacks, and of course general I have created a detailed map of use cases.

But it works because the stakes are low. This is similar to how we perceive advances in robotics. In other words, if the robot does a backflip, it will be surprised, and if it falls while trying to open the drawer, it will be fine. We wouldn’t be so philanthropic if it had dropped test bottles in hospitals. Luckily it wasn’t.

Joined Microsoft. (And Google, for that matter, just rushed to play while Microsoft was fervently pursuing its own goals.)

Microsoft made a big mistake. In fact, it’s Bing’s mistake.

Last week’s big announcement made custom BingGPT (not what they called it, but they use to disambiguate in the absence of a sensible official name) safer, smarter, and more capable. . In fact, there was a whole special wrapper system called Prometheus that seemed to mitigate the chance of bad responses.

Unfortunately, as anyone familiar with Pride and Greek mythology could have predicted, it seems that Prometheus endlessly skipped straight to the part where he’s very openly having his liver ripped out.

Oops, AI did it again

To begin with, Microsoft made the strategic mistake of tying its own brand too much with OpenAI’s brand. As an investor and stakeholder in the investigations GPT is conducting, GPT has been removed and not accused of any hoax perpetrated by GPT. But someone made the heady decision to go all in on Microsoft’s already somewhat dodgy branding of Bing, turning conversational AI’s worst tendencies from curiosity to liability.

As a research program, ChatGPT is allowed a lot. However, as a product, it used to be largely unreliable, with claims on the box like how it can help you write reports, plan trips, and summarize the latest news. Trust no one now. Even what must have been the best-case scenario that Microsoft laid out in its own presentation of the new Bing was riddled with errors.

These errors are not caused by OpenAI or ChatGPT. With Microsoft’s decision to own the messaging, branding, and interface, anything that goes wrong becomes Bing’s problem. And to Microsoft’s further misfortune, Microsoft’s perpetually inferior search engine becomes like the bastard’s indiscretion in the old jokes. No, it’s not. ” One failure means eternal skepticism.

One failed trip upstate means no one trusts Bing to plan vacations. His one misleading (or defensive) summary of a news article means that no one believes it to be objective. His one-time repetition of vaccine disinformation means trusting no one to know if it’s real or fake.

And Microsoft has already nudged that this won’t be a problem thanks to Prometheus and the “next generation” AI it manages, so even if Microsoft says “fixed!” do not trust

Microsoft poisoned the well they just put Bing in. Today, consumer behavior is fickle and its consequences are difficult to predict. This surge of activity and curiosity will probably make some users stick around, and even if Microsoft delays a full rollout (and I think it will), the net effect will be an increase in Bing users. A victory for Pyrrhus, but a victory nonetheless.

What worries me more is that tactical Microsoft didn’t seem to understand what technology it considered suitable for commercialization and promotion.

“Ship it as is.” -Someone, probably

On the very day that BingGPT was first demonstrated, my colleague Frederic Lardinois very easily managed to get consumer AI to do two things it shouldn’t. No cautions or warnings.

Large AI models clearly have a fractal attack surface, cleverly improvising new weaknesses where old ones have been strengthened. People use it all the time. In fact, it’s good for society, and more recently he’s also for OpenAI, to show how an avid quick hacker can circumvent safety systems.

It would be kind of terrifying if Microsoft decided to be peaceful with the idea that someone else’s AI model with a Bing sticker on it would be attacked from every quarter and likely to say some really weird things. . Risky, but honest. Just say it’s a beta like everyone else.

But they didn’t seem to realize this would happen.In fact, they don’t seem to understand the nature or complexity of the threat at all. And this is after Tay’s infamous corruption! Of all companies, Microsoft should be the most cautious about releasing simple models that learn from conversations.

Some would argue that before betting on any significant brand (in that Bing is Microsoft’s only bulwark against Google in search) some testing is required. The fact that BingGPT has all these nasty issues in his first week since it came out seems to prove that Microsoft doesn’t test BingGPT properly internally. It can fail in many ways, so details can be left out, but the end result is incontrovertible. The new Bing was simply not ready for general use.

This now seems obvious to everyone in the world. Why wasn’t that revealed to Microsoft? Perhaps they were blinded by the ChatGPT hype and, like Google, decided to rush ahead and “rethink search”.

People are rethinking search now. They’re rethinking whether either Microsoft or Google can be trusted to deliver virtually correct search results on a fundamental level, AI-generated or not! (Neither Meta) has demonstrated this ability at all, and the few other companies working on this challenge have yet to do so at scale.

I don’t know how Microsoft can save this situation. Using their relationship with OpenAI to leapfrog stupid Google, they committed to the promise of new Bing and AI-powered search. They can’t bake cakes.

It is very unlikely that they will withdraw completely. It would be a massive embarrassment – even more than we’re experiencing right now.

Likewise, it’s hard to imagine Microsoft rushing in as if nothing was wrong. That AI is really weird! Sure, they’re forced to do this, but they’re making threats, claiming multiple identities, insulting users, and hallucinating all over the place. They must admit that their allegations of inappropriate behavior controlled by poor Prometheus are, if not lies, at least untrue. You obviously haven’t tested this system properly.

The only reasonable option for Microsoft is the one they seem to have already adopted. That means suppressing invites to the “new Bing” and kicking the can to release a few specific features at once. Maybe even give the current version an expiration date or a limited number of tokens so that the train eventually slows down and stops.

This is the result of deploying a technology that I didn’t invent, didn’t fully understand, and didn’t fully appreciate. This debacle could have set back major deployments of AI in consumer applications. This probably suits him for OpenAI and others building next generation models.

AI may be the future of search, but it’s certainly not the present. Microsoft chose a very painful way to find it.