Bing around and find out • TechCrunch

The new and the Improved Bing, powered by a custom version of OpenAI’s ChatGPT, had an incredibly fast reversal: from “next big thing” to “brand-sinking albatross” in less than a week. Well, it’s all Microsoft’s fault.

ChatGPT is a really cool demonstration of new and unknown technology that’s also fun to use. So it’s no surprise that, like every other AI-adjacent construct that comes along, this novelty would cause its abilities to be overestimated by everyone from high-tech types to people normally indifferent to space.

It’s at the right “technology readiness level” to discuss over tea or beer: what are the merits and risks of the generative AI approach on art, literature or philosophy ? How to be sure that it is original, imitative, hallucinated? What are the implications for creators, coders, customer service reps? Finally, after two years in crypto, something interesting to tell!

The hype seems overblown partly because it’s a technology more or less designed to provoke discussion, and partly because it borrows from the controversy common to all advances in AI. It’s almost like “The Dress” in that it orders a response, and this response generates other responses. The hype is itself, in a way, generated.

Beyond simple chat, large language models like ChatGPT are also well suited for low-stakes experiments, e.g. Endless Mario. In fact, that’s really OpenAI’s fundamental approach to development: first release models privately to get the sharpest edges out of them, then publicly to see how they react to a million people kicking the tires simultaneously. At some point, people give you money.

Nothing to gain, nothing to lose

What is important in this approach is that “failure” has no real negative consequences, only positive consequences. Calling its models experimental, even academic in nature, any participation or engagement with the GPT model series is simply a large-scale test.

If someone builds something cool, it reinforces the idea that these models are promising; if someone finds a significant failure state, well, what else did you expect from an experimental AI in the wild? He sinks into darkness. Nothing is unexpected if everything is – the miracle is that the model works as well as it does, so we are perpetually satisfied and never disappointed.

In this way, OpenAI has harvested an astonishing volume of proprietary test data with which to refine its models. Millions of people testing GPT-2, GPT-3, ChatGPT, DALL-E and DALL-E 2 (among others) have produced detailed maps of their capabilities, shortcomings and, of course, use cases popular.

But that only works because the stakes are low. It’s similar to how we perceive advances in robotics: amazed when a robot does a backflip, indifferent when it falls trying to open a drawer. If he dropped test vials in a hospital, we wouldn’t be so charitable. Or, for that matter, if OpenAI had loudly asserted the security and advanced capabilities of the models, but thankfully, that wasn’t the case.

Enter Microsoft. (And Google, for that matter, but Google just rushed the game while Microsoft diligently pursues its own goal.)

Microsoft made a big mistake. A Bing error, actually.

His big announcement last week was quick to make claims about how he worked to make his BingGPT custom (not what they called it, but we’ll use that as a disambiguation in the absence of sensible official names) safer, smarter and more capable. In fact, he had a whole special wrapping system called Prometheus that supposedly mitigated the possibility of inappropriate responses.

Unfortunately, as anyone familiar with pride and Greek myth could have predicted, we seem to have skipped straight to the part where Prometheus continually and very publicly gets his liver ripped out.

Oops, the AI ​​did it again

First, Microsoft made a strategic mistake by linking its brand too closely to that of OpenAI. As an investor and interested party in the team’s research, she was out of the way and blameless for any shenanigans GPT engages in. But someone made the far-fetched decision to push ahead with Microsoft’s already somewhat laughable Bing brand, converting conversational AI’s worst trends from curiosity to responsibility.

As a research program, a lot can be forgiven for ChatGPT. As a product, however, with claims on the box like how it can help you write a report, plan a trip, or summarize recent news, few people would have trusted it before and no one will now. Even what were supposed to be best-case scenarios released by Microsoft in its own presentation of the new Bing was riddled with errors.

These errors will not be attributed to OpenAI or ChatGPT. Due to Microsoft’s decision to own the messaging, branding, and interface, anything that goes wrong will be a Bing issue. And it’s another misfortune of Microsoft that its eternally outdated search engine will now be like the barnyard indiscretion of the guy in the old joke – “I built this wall, do they call me Bing the mason? No, they don’t. Failure means eternal skepticism.

A failed trip upstate means no one will ever trust Bing to plan their vacation. A misleading (or defensive) summary of a news article means that no one will believe it can be objective. A repetition of vaccine misinformation means that no one will trust him to know what is true or false.

Prompt and response to Bing’s new conversational search.

And since Microsoft has already sworn that won’t be a problem thanks to Prometheus and the “next-gen” AI it powers, no one will trust Microsoft when they say “we fixed it!” »

Microsoft has poisoned the well it just threw Bing into. However, the vagaries of consumer behavior are such that it is not easy to predict the consequences. With this spike in activity and curiosity, perhaps some users will stick around and even if Microsoft delays the full rollout (and I suspect they will), the net effect will be an increase in Bing users . A Pyrrhic victory, but a victory nonetheless.

What worries me the most is tactical mistake made by Microsoft in apparently not understanding the technology it saw fit to produce and evangelize.

“Just ship it.” -Someone, probably

On the very day that BingGPT was first demonstrated, my colleague Frédéric Lardinois managed, quite easily, to get it to do two things that no mainstream AI should do: write a hateful screed from Adolf’s point of view Hitler and offer the aforementioned vaccine misinformation with no caveats or warnings.

Clearly, any great AI model presents a fractal attack surface, sneakily improvising new weaknesses where old ones are reinforced. People will always take advantage of this, and in fact, it’s in the interests of society, and lately, in the interests of OpenAI, that dedicated hackers will show ways to circumvent security systems.

It would be kind of scary if Microsoft decided it was okay with the idea that someone else’s AI model, with a Bing sticker on it, would be attacked from all sides and probably say stuff really weird. Risky, but honest. Say it’s a beta, like everyone else.

But it really seems like they didn’t realize it was going to happen. In fact, it seems they don’t understand the character or complexity of the threat at all. And that’s after Tay’s infamous corruption! Of all the companies, Microsoft should be the most hesitant to release a naïve model that learns from its conversations.

You would think that before betting on a major brand (as Bing is Microsoft’s only bulwark against Google in search), a number of tests would be necessary. The fact that all of these troubling issues surfaced within BingGPT’s first week of existence seems to prove beyond doubt that Microsoft hasn’t tested it enough internally. It could have failed in various ways so we could ignore the specifics, but the end result is indisputable: the new Bing just wasn’t ready for general use.

It seems obvious to everyone now; why was it not obvious to Microsoft? Presumably, he was blindsided by the hype for ChatGPT and, like Google, decided to swoop in and “rethink search”.

People are rethinking search now, okay! They are rethinking whether Microsoft or Google can be trusted to deliver search results, AI-generated or not, that are even factually correct at a basic level! No company (nor Meta) has demonstrated this capability, and the few other companies taking up the challenge have yet to do so at scale.

I don’t see how Microsoft can save this situation. In an effort to leverage their relationship with OpenAI and move past a hesitant Google, they committed to the new Bing and the promise of AI-powered search. They can’t undo the cake.

It is highly unlikely that they will pull out completely. It would involve full-scale embarrassment – ​​even greater than he is currently experiencing. And because the damage is already done, it might not even help Bing.

Likewise, it’s hard to imagine Microsoft loading as if nothing had happened. His AI is really weird! Sure, it’s coerced into doing a lot of these things, but it’s uttering threats, claiming multiple identities, shaming its users, hallucinating everywhere. They have to admit that their claims about inappropriate behavior controlled by poor Prometheus were, if not lies, at least not true. Because as we have seen, they clearly did not test this system properly.

The only reasonable option for Microsoft is the one I suspect it has already taken: speeding up the invites to the “new Bing” and launching the box on the road, unleashing a handful of specific features at once. Maybe even give the current version an expiration date or a limited number of tokens so the train will eventually slow down and stop.

It’s the result of deploying technology that you didn’t originate, fully understand, and can’t satisfactorily assess. It’s possible that this debacle has delayed major deployments of AI in consumer applications by a significant period of time – which is probably fine with OpenAI and others who are building the next generation of models very well.

AI may be the future of research, but it is certainly not the present. Microsoft chose a remarkably painful way to find out.