Google vs Microsoft (Part 1): Microsoft’s New Bing Is a Paradigm Change for Search and the Browser
The AI cold war between Google and Microsoft is over
This is part 1 of a three-part article series covering the recent news on AI involving OpenAI, Microsoft, Google, ChatGPT, the New Bing, and Bard, and how the events will unfold into a new era for AI, search, and the web (part 2 & part 3).
A frantic week. That’s the best way to describe what the AI industry has lived since Monday. Google and Microsoft are at the hottest point in their long-term relationship and everything is because of ChatGPT. In a surprise event on Tuesday, Microsoft announced the New Bing, powered by a “next-generation” AI model by OpenAI, just one day after Google disclosed they will soon release a ChatGPT competitor named Bard, and one day before its official presentation. Perfect timing.
We’re living through a paradigm change in the way companies do AI with OpenAI, Microsoft, and Google (among others) as the leading characters of what promises to be an epic play. AI has turned into a product-focused blooming landscape that keeps the foot on the gas pedal. Many predicted this outcome the very moment OpenAI released ChatGPT—which recently topped off a record-breaking ascent to being the “fastest-growing consumer application” ever.
Even zooming out, these events—from ChatGPT’s release on Nov 30 to Microsoft’s and Google’s announcements this week—reflect a salient milestone in the history of AI (even if not technically speaking, at least at every other level). The AI industry will be deeply marked by them for the foreseeable future.
This is my attempt to frame the facts objectively and hypothesize on the impending future that’s looming over us. I tried to keep this short but eventually realized I’d have had to cut too much info. Instead, I’ve decided to divide what was initially a shy 2000-word essay, into a three-part long series that I’ll be publishing in the coming days.
This first part is all about Microsoft (kind of). How the company wants to redefine software with the help of its partner, OpenAI, and its beloved ChatGPT, and how they plan to fight Google in what could be the fiercest tech battle of the century—for the dominance of search, the web, and AI. It’s divided into five subsections:
The most successful synergy in the AI landscape
Bing + Edge: Your AI copilot for the web
A next-gen ChatGPT and the Prometheus model
With great power comes great responsibility
What does Microsoft really want?
The most successful synergy in the AI landscape
“It’s a new day in search,” announced Satya Nadella, Microsoft Chairman and CEO, “a new paradigm for search.” He revealed on Tuesday’s surprise event the company’s intention to reimagine the “largest software category” in the world. Microsoft will blend together the search engine and the web browser as the two pillars of an “AI copilot for the web”—a remix of Bing (“the new Bing”) and Edge, combined with chat capabilities powered by a “next-generation large language model” from OpenAI “more powerful than ChatGPT.”
So much ambition in one paragraph—an ambition that this week solidified into a tipping point moment, the outcome of a brilliant strategy that Nadella initiated four years ago, in 2019, with a $1B investment into the young AI startup led by Sam Altman. OpenAI, the mother of the now uber-popular GPT models was at the time merely planting the seed (GPT-2) that eventually blossomed into ChatGPT.
Nadella’s visionary move changed the paths of both companies forever and, if his plans come to fruition, will be considered among the greatest business decisions in the tech space during the last decade. After doubling down with $10B more last month, Microsoft was ready to take the next step and try to overcome a challenge that has, for more than two decades, proved unapproachable: Dethrone Google from its unending hegemony in both the AI and search spaces.
“The race starts today … and we’re going to move fast,” Nadella said. “I hope that with our innovation [Google] will definitely want to come out and show that they can dance and I want people to know that we made them dance.” Shots fired.
But enough storytelling, let’s get to the facts. What has Microsoft unveiled this week? Is it really deserving of such attention and ovation or is it just a good narrative design? Is Bing a real threat to Google? Can these AI-enhanced products actually work as a viable business?
Here’s what we know about the new Bing, the changes to Edge, and how they come together into what Nadella anticipates is going to be a key stepping stone toward the third revolution of the internet.
Bing + Edge: Your AI copilot for the web
Everything Microsoft announced on Tuesday stems from their motivation to reimagine how we interact with the web by means of the latest generative AI technology, provided by OpenAI. They’ve upgraded the core search algorithm so that “basic search queries are more accurate and more relevant” and enhanced two of the company’s primary software products—the Bing search engine and the Edge web browser—that together comprise, or so they hope, what will be our future AI web copilot (in the line of GitHub Copilot but for searching, chatting, and writing).
The new Bing
The new Bing (that’s the actual name—Nadella, in a striking display of trust, rejected the opportunity to rebrand it because “brands can be rebuilt as long as there’s innovation,” as he told Nilay Patel, The Verge’s editor-in-chief in an exclusive interview) embodies the biggest shift that search, as a software category, has experienced in more than two decades (Google Search makes 25 this year).
Yusuf Mehdi, Corporate VP and Consumer CMO at Microsoft, recalled on Tuesday that “the user experience … [is] essentially the same as 20 years ago.” The technology behind search has changed gradually over the years but incremental changes don’t take a product to the next level. Yusuf explains that search works well only for about 50% of queries (navigational, e.g. “go to Substack.com” and informational, e.g. “what time is it in San Francisco?”), which means “roughly half of all searches aren’t delivering the job that people want” (e.g. “trip planning”).
The new Bing is divided into two differentiated applications. First, the search engine as we know it. It provides links as usual but it’s now combined with an assistant that gives direct answers to queries, including a list of links from which it takes (at least part of) the information it generates. The search box now accepts up to 1,000 characters to make room for long queries.
This feature is available as a preview on desktop (soon on mobile) and will be rolled out to “millions in the coming weeks.” You can sign up here to get on the waitlist.
The second application is a built-in chat interface similar to ChatGPT with the additional ability to cite sources. It’s better suited for complex searches (that 50% of queries where traditional search engines fall short).
Here are more examples. Karen X. Cheng, who followed the event in person, got access to the new Bing and offered her followers to submit queries, advising them to “try things that search engines today aren’t good at. Instead, treat it like a personal assistant.” That’s where these products allegedly shine. Tom Warren, senior editor at The Verge, who covered the event, run some more queries on Twitter:


At a first glance, you can appreciate that this doesn’t look like an incremental improvement. It’s something bigger—like what Neeva.com, Perplexity.ai, and You.com have been promoting for weeks—substantially different than what Bing was before.
It’s not just a step up in terms of the technology that powers the product (search engines have used AI algorithms for years, but the qualitative contrast now is notable), but also in terms of the UI/UX, which is generally accepted as one of the core reasons why ChatGPT was a tremendous success despite being “nothing more” than a slightly improved version of a 2yo piece of tech. (As a side note, in an interview with Alex Konrad for Forbes, Sam Altman revealed that he “pushed hard for [ChatGPT]” despite the team’s reservations toward the launch. “I really thought it was gonna work,” he said. That’s vision.)
Despite the improved UX and functionality, I’ve read people wonder why would anyone swap to Bing when you can simply use Google Search and ChatGPT as separate tools for different tasks (we shouldn’t dismiss this idea right away, more on this later). I’ll get deeper into the underlying technology that powers the product soon but for now, I’ll say that the new Bing and ChatGPT are simply incomparable tools (not to say one is better than the other for any given task—they’re simply different): search capabilities, up-to-date information, and enhanced performance are some of the reasons why.
As James Vincent writes for The Verge, “unlike ChatGPT, the new Bing can also retrieve news about recent events … the search engine was even able to answer questions about its own launch, citing stories published by news sites in the last hour.” Despite claims that ChatGPT struggles with this, the truth is OpenAI never wanted users to access the internet with ChatGPT—the capability was always there (even if in prototype form) as people showed after successfully “jailbreaking” the model to reveal the instructions set by OpenAI researchers (e.g. “Browsing: disabled”).
Edge web browser
The second product Microsoft unveiled is an AI-enhanced Edge web browser (as I interpret it, the AI copilot is the combination of the entire suite of products; search, answers, chat, and browser, but I’ve seen others refer to the new browser as the AI copilot, just to avoid any possible confusion). Is the new Edge—not search—what excites Greg Brockman, OpenAI’s President (his tweet shows the interface of the AI assistant in the browser):



They took all the capabilities of the new Bing and installed a sidebar in the browser that allows the user to access those features without the need to use Bing (it’s a smart move by Microsoft to dissociate the new AI features from any particular product given that pretty much no one uses Bing—for now).
Also, Nadella told Patel that the product could be easily implemented on Google Chrome. When Mehdi was asked during the Q&A whether they’d allow it to be on Chrome, he said that, although they’re starting with Edge, “our intention is to bring it to all browsers.” Another intelligent decision to reduce friction.
On the new browser, the underlying AI model can interact with the websites you visit, mainly through two features, chat and compose. Chat, as you may have guessed, allows you to access the chatbot that powers Bing (e.g. to summarize a 15-page PDF document). Compose acts as a writing assistant, pretty much like any other writing tool that uses state-of-the-art language models (LMs).
The new Bing and the AI-enhanced Edge embody Microsoft’s attempt at reimagining search and our overall experience on the web. But how does this work? How did they manage to transform a pure chatbot like ChatGPT into what they see as the next revolution of the internet?
A next-gen ChatGPT and the Prometheus model
OpenAI’s “next-generation large language model” is the answer. Microsoft’s products sit comfortably on top of OpenAI’s nameless AI model, which is “more powerful than ChatGPT” and it’s already tailored for search. But that’s not everything. In between, there’s a mix of “capabilities and techniques” that Microsoft uses to further improve the AI’s adequacy for search. They call it the Prometheus model. That’s what powers the new Bing and Edge’s capabilities.
Through Prometheus, Microsoft has found a way to bring together the generative capabilities of reinforced LMs like ChatGPT, with the reliable retrieval skills of a search engine (the resulting system may not be as reliable, though, as we’ll soon see). “We’ve done a lot to the [Prometheus] model to ground it in search,” Nadella told Patel. Microsoft claims that Prometheus provides “more relevant, timely and targeted results, with improved safety.”
Although Microsoft hasn’t disclosed the specificities of what they’ve done to transform a chatbot like ChatGPT into Prometheus, which can generate direct answers to search queries, cite sources, and retrieve up-to-date information—it’s pretty clear that, as I advanced above, they’re hardly comparable. They’re not equal either in terms of performance or in terms of design. The Prometheus model is simply a different tool, tailored for a different task at the engineering and design levels (in any case, the degree to which it is really a tool appropriate for search is for users to decide).
It’s paramount to frame adequately how to use tools whose internal structure and function we ignore (like ChatGPT or the new Bing). In one of my latest articles, “5 Practical Applications Where ChatGPT Shines,” I underscored the importance of using ChatGPT right (even if there’s no intentional design in that case). We have to learn how the tech works and understand the boundaries that separate the applications for which ChatGPT is a good tool and those for which it isn’t.
Base ChatGPT is not suitable for search. Neither are pure LMs like GPT-3 or Chinchilla, nor chatbots like LaMDA and Sparrow. It’s not trivial to transform ChatGPT into a tool for search and it’s yet unclear whether Microsoft has successfully achieved the desired result. My prediction, not having tested it personally yet, is that despite being “next-gen” it will fail. And it will fail in pretty much the same aspects ChatGPT fails (maybe not to the same degree)—with the added drawback that now we’re not talking about a “research preview”, but about a product from a well-known megacorporation.
As a side note, you may be wondering whether this “super ChatGPT” is actually GPT-4. I don’t know. It seems GPT-4 was ready to be released but OpenAI prioritized ChatGPT. Altman referred to the model simply as the “next-generation model” and Nadella—after Patel tried to get the scoop with a teasing “is it GPT-4?”—only said: “Let Sam [Altman] at the right time talk about his numbers.” As far as I know, no one from Microsoft or OpenAI has confirmed or denied anything.
With great power comes great responsibility
There’s a lot to unpack here. Microsoft’s new products promise to be an amazing opportunity for the company to close the gap with Google, which, at least until now, dominated those areas without much competition. However, revolutions and radical changes bring a lot of repercussions. Some can be prevented but others are simply unforeseeable. Let’s explore the problems and consequences these products could create for the space and the users. I’ll classify them roughly into two levels, product/application (Bing and Edge) and tech/model (next-gen ChatGPT and Prometheus).
Product/application level
How will Bing and Edge impact users and the landscape in general? I’ll draw examples and questions from Patel’s interview as well as from the Q&A that Microsoft execs conducted after the event.
Giving value back to creators
One of the core business aspects of search is the reciprocal relationship between the owners of websites (content creators and publishers) and the owners of the search engines (e.g. Google and Microsoft). The relationship is based on what Nadella refers to as “fair use.” Website owners provide search engines with content and the engines give back in form of traffic (or maybe revenue, etc.). Also, search engine owners run ads to extract some profit from the service while keeping it free for the user (a business model that Google popularized and on top of which it amassed a fortune).
Patel asked Nadella a few times about how the new products could affect this symbiotic relationship. In the new Bing, you can simply ask the AI assistant to give you the best 10 healthy recipes for lunch that contain chicken and avocado, and instead of clicking on the web links, you may decide to read the direct answer, without ever entering the website from which the assistant took the info (even if the sources are there).
Nadella answered Patel’s question rather tangentially. He said that they care and keep in mind that the only way to make this work is to give back to the owners of the websites. “Everything you saw there had annotations, everything was linkable,” he said. “This is just a different way to represent the ten blue links.” When Patel reformulated the question after what felt like an unsatisfactory answer, Nadella told him that “even search today … has answers … I don’t think of this as a complete departure from what is expected of a search engine today.” As I interpret it, Nadella wants to believe people won’t take the assistant’s answers at face value and will instead go to the website to contrast the info.
Data tells a different story. On the one hand, there’s anecdotal evidence that people have been using ChatGPT as a source of information or as a customized teacher—even though the chatbot isn’t directly designed for this, unlike the new Bing. On the other hand, Sridhar Ramaswamy, ex-Google SVP and founder of Neeva (a direct competitor of Bing and Google Search), says that “as search engines become answer engines, referral traffic will drop! It’s happened before: Google featured snippets caused this on 10-20% of queries in the past.”
The bottom line is that people trust these AI models (aggravated by the inevitable automation bias) and, at the same time, they won’t contrast the information as often as website creators would like to. Ramaswamy says that “LLMs and AI answer bots aggregating and summarizing [content]” will make things “a whole lot worse for [publishers].” The way to solve the problem, as Ramaswamy suggests, is through a healthier relationship between search engine owners and publishers.
The threat of AI-generated content
Something that grabbed my attention throughout the event and the interviews is Microsoft’s repeated emphasis on the concept of “Copilot”. As I see it, they want to convey the idea that the new products (search, browser, chatbots, etc.) aren’t disconnected from the person that uses them. They want to convince us that there’s always a human in the loop—in their view, that’s the person who prompts the AIs, either with a search query, a direct question, or with the intention to write a creative fiction story. They want to emphasize that human and AI are inseparable.
This is a discursive defense from criticisms of the form “AI will replace people or take our jobs” or “what if AI-generated content floods the internet.” The copilot concept blurs the boundary between AI-generated content and original, human-made content. Karen X Cheng asked during the Q&A if things will be marked as AI-generated to which Yusuf Mehdi said that “our vision is copilot … we don't want Bing to write things completely.” Spoiler: no.
I wrote an essay on the topic a while ago where I argued that AI-generated content flooding the web could be a serious issue, maybe not for the data that the models consume but for the content we consume. We can’t predict what will happen if a product that has the potential to create instantaneous blog posts reaches hundreds of millions of users as Microsoft would like (I know I’m being generous… we’re talking about Bing, but there’s an implicit intention to pressure Google to follow suit), but we may find out soon.
Monetization and cost structure
Search will change but, for now, not the way it’s monetized. Microsoft will keep the ad business model from the start (the service will remain free for both the new Bing and the Edge browser). The question is then, can Microsoft really afford to run these products with ads as the only source of revenue when the queries cost a lot more for LMs than traditional search engines? Mehdi didn’t answer Fortune’s Jessica Mathews’ question about the cost of running a query on GPT vs the old Bing, so maybe that’s the answer (in case you want to know more, Dylan Patel wrote a great essay for Semianalysis on the topic).
But maybe the question is even more tricky: Does Microsoft really intend to run these products at a profit? There’s a lot to gain here besides revenue. Microsoft is an extremely large, powerful, and diversified company. They could afford to run this new Bing at a loss for some time in an attempt to choke Google or pressure them to react faster.
Tech/model level
We know about this very well because I’ve written once and again for TAB about this. Biases, hallucination, unsafe filters, prompt injection… all the typical flaws that pain modern LMs like GPT-3, ChatGPT, or LaMDA, are likely to be present in Microsoft’s new products. Whatever they’ve done with Prometheus to improve the performance and search skills of the products, is unlikely to solve these issues. The reason is that they emerge from how the underlying technology (i.e. the “super ChatGPT”) is designed and trained—and that hasn’t changed.
Not unlike OpenAI, Microsoft adheres to a set of AI principles that Nadella says aren’t just words on a document—they’ve “been practicing” them. The excuse to go out into the world and present a product that may not be completely ready for prime time is that, in the words of Nadella, AI is “about alignment with human preferences and societal norms. You're not going to do that in a lab.”
I agree that feedback from real users is highly valuable, as the folks at OpenAI have corroborated repeatedly, but a conflict arises when you have to trade off caution for boldness—it’s not always easy to find the sweet spot. That’s the main point where Microsft and Google differ. Microsoft’s stance aligns much better with OpenAI’s. They want this technology to be out in the world even if imperfect. They can always find ways to improve it retrospectively.
Loyal to this viewpoint and convinced of its ability to adhere to those responsible principles, Microsoft will launch the new Bing to millions soon. Sarah Bird, responsible AI Lead at the company said in the Tuesday event that “we’ve gone further than we ever have before to develop approaches to measurement to risk mitigation,” and, to a question about the accuracy of the new Bing, Dena Saunders said that “we're not going to always get it right … the key here is really how we ground information.”
But those are just words. What users have found unsurprisingly contrasts with the enthusiastic tone of the event and the positive reactions of the press and people across social media. It’s funny that a factual error (that LMs make all the time) botched Google’s Wednesday Bard demo (to the point of making the company lose $100B) only for Bing to commit the exact same errors as soon as users could put their hands on the product.
I’m going to focus mainly on hallucinations and a bit on jailbreaking, but I’ve found also instances of bias and discrimination.
Hallucinations (aka making stuff up)
LMs make up information. It doesn’t matter if they’re enhanced with RLHF or grounded on search with undisclosed techniques. Anything you put on top can theoretically be bypassed so that the model fails to do the task correctly (sometimes without the user realizing it).
Harry McCracken for Fast Company asked during the Q&A if they had solved this problem. Bird said that “we’ve been working on this since the beginning [but] it's not perfect, users will see where it's changing a small number or something.” Given the kinds of hallucinations that we all have seen ChatGPT make, saying that it may change “a small number or something” sounds like an attempt to underestimate the importance of this—much more when you’re selling the product as a substitute for the traditional search engine, the main tool to retrieve information safely and reliably from the web.
I didn’t have to search for too long to find an example. Gael Breton posted a thread on Thursday: “content creators & SEOs will find both terrifying and relieving.” A group of Reddit users with access to the new Bing tested it and found two things: First, it truly is “ChatGPT on steroids,” and second:


You may think, “well, yeah, ChatGPT also makes stuff up, I’m not surprised.” But we should go deeper than that. Yes, ChatGPT is a bullshit generator. People who use it as a source of information are using it wrongly. If people use it that way is because it seems to be able to do the task but it wasn’t designed for that and OpenAI never claimed it was. The responsibility lies with the user. But the new Bing is different. This is a product “grounded on search,” tailored at the model and application levels. It's designed to do the task and presented as a tool that can do the task. But it can’t, at least not with the high reliability that characterizes the traditional search engine.
This isn’t surprising to me (and shouldn’t be to you). In the end, the core tech that powers the product is essentially the same as ChatGPT; why would anyone expect that band-aids could prevent the product from showing its true nature? As Gary Marcus says, “hallucinations are in their silicon blood, a byproduct of the way they compress their inputs, losing track of factual relations in the process.”
But there's a set of bigger problems that didn’t exist with ChatGPT. First, people will, rightfully, believe Bing's responses much more than they believe ChatGPT's. Microsoft is succeeding in selling the story that this is the next generation of the search engine—a reimagination of how we interact with the internet (and we use the internet to look up info). If search was reliable before, it’ll be after, people will think. But that's not the case. Now, people who have been warning all of us about this have to combat Microsoft's word in addition to people's unwillingness to think critically.
Second, this is a product by Microsoft. Let's not forget that (and let's set aside for a moment that we’re talking about Bing because we risk freeing Microsoft from accountability just because the original brand is horrible). ChatGPT has been a success story worthy of AI history books, but so far it has reached, in the best estimates, 100M users (the NYT says it’s 30M). Although undoubtedly impressive, it's not close to the 1B+ number that big tech companies like Microsoft and Google play with. The new Bing is an unreliable product presented as more reliable than ChatGPT and with the potential to reach many more people. Great.
And three, we don’t know exactly how Bing merges search with chat. How it modifies the information from the sources it takes. When it shows verbatim answers and when it relies on its generative capabilities. Let me ask you this: What is worse, knowing that you shouldn’t trust ChatGPT because everyone admits it makes up stuff or living in the uncertainty of not knowing if you should trust your brand new search engine to do a task it’s supposedly designed to do?
Of course, users can always go check the sources, but then, why put the generative bit there in the first place?
Jailbreaking (aka bypassing filters)
People successfully jailbroke ChatGPT out of its guardrails, carefully set by OpenAI to avoid the AI from misbehaving, the day after launch. On Dec 1, Riley Goodside, expert prompter if there’s one, posted this:


People have successfully “jailbroken” the new Bing one day after launch, too:
People found a codename, “Sidney”, hiding somewhere in the underlying chatbot. It isn’t a problematic jailbreak like the one called DAN (Do Anything Now) that’s been circulating on Reddit, but it’s a hint at existing vulnerabilities. Every time a new jailbreak for ChatGPT appeared, OpenAI engineers patched it. But that’s not how this should work because, eventually, it may cause serious harm before it’s solved.
What does Microsoft really want?
This essay is already super long, but we should try to make sense of it all in the big picture. The question that will help us connect this article with the remaining two is this: What does Microsoft really want?
As I see it, this masterful move by Microsoft has two motivations (maybe just one of them or some weighted combination). First, Microsoft wants to leverage the opportunity that generative LMs—and its partnership with OpenAI—bring to transform the “hobby” that is search right now (profitable but “negligible” in terms of revenue and market share ~4%) into a main product for the company. They may intend to increase revenue, take a good chunk of market share from Google (which is at 80%+), or even expand the product category to attract new users. Nadella says that generative AI entails “a tremendous opportunity for us to make some real progress here.”
Second, they may be looking for a way to pressure Google without a real intention to take a significant number of customers or revenue from them. Taking Google out of its comfortable spot at the top on both AI and search may just require a light push to the throne chair on Microsoft’s part. As Nadella said, they want to make Google come out and dance. And in between precise foot movements and carefully designed flourishes, they may attack the vulnerabilities to turn the tides upside down.
I’ll expand much more on this analysis and what it’ll mean for the AI industry at large in the second and third chapters of this series. For now, I’ll follow François Chollet's advice (although he’s obviously biased in favor of Google) and take a level-headed approach toward everything that Microsoft has unveiled this week. I focused on the problems not because I consider the new Bing a useless product (I actually think it could prove tremendously helpful), but because I see hype skyrocket after every event or demo—and that’s simply a bad mindset to analyze what’s happening.
I leave you now with Chollet’s words:
“Don't judge the performance of an AI system based on specific examples. 100% of the value of an AI system comes from its ability to generalize broadly, which you cannot see through a few specific examples.”
I wonder how did Msft incorporate LLM into the search
Do they create or use another model ground-up that can do traditional search and the stuff that chat-gpt could?
Or they perform a traditional search and using the content of the search results to answer like a chat-got
Or some other ways?
Ultimately, the function of search , I would believe, is to find all relevant, reliable information. It isn’t to process or analyse information. Chat-gpt seems to incline more towards that. So perhaps its better to use these 2 separate , especially when one do not have enough subject matter knowledge.
When Sam came on stage he said that the new model was faster, more accurate and more capable than ChatGPT but based on GPT 3.5 and the learnings they did with the ChatGPT Research preview.
I look at this like more like a "Distilled ChatGPT" connected to the internet than a "sparse ultra large LLM" like GPT4 is supposed to be.
That would make sense because if they really want to scale this tech they need to reduce the compute necessary.
If the next model is so compute intensive that it is more in dollars per query than cents per query, I believe they will keep their flagship model to themselves and make it available via ChatGPT-Pro or ChatGPT-Plus to make up for the cost.
Having access to superhuman AI at all would be such a game changer that I still have a hard time truly grasping what that would mean. I guess we will find out soon enough.