The Fuck-Ups of the Company That Invented the Technology

Google invented the transformer. That is not some minor historical footnote buried inside an old research paper. The transformer architecture is the foundation on which almost every serious modern AI model was built. OpenAI took it, scaled it aggressively, put it inside a product that ordinary people could actually use, and lit the entire AI boom with GPT. Google created the core technology and then spent the next several years finding increasingly creative ways to fuck up the execution.

That does not mean Google lacks talent. Quite the opposite. Google has some of the best researchers, engineers, infrastructure and distribution in the world. It owns Chrome, Android, Search, Gmail, YouTube, Docs and one of the largest cloud platforms on Earth. It has more opportunities to introduce useful AI into ordinary people’s lives than almost any other company, which is precisely why its repeated failures are so difficult to understand. Google is not being held back by a lack of money, researchers, computing power or access to users. It is being held back by its apparently endless inability to turn those advantages into coherent products. What the fuck are they doing?

From Bard to Gemini 3: the brief moment they looked competent

Bard was a disaster of a name and a mediocre launch. It felt rushed, defensive and fundamentally unsure of what it wanted to be. Google had spent years developing much of the technology behind modern language models, only to appear completely surprised when ChatGPT turned that technology into a product people actually wanted. The Gemini rebrand was an improvement, and the models from Gemini 1.5 through 2.5 were generally fine. They were not consistently the best models available, but they were useful, reasonably fast and often competitively priced. Their enormous context windows were genuinely valuable, while the Flash models were cheap enough that people could tolerate their weaknesses.

Google also made the sensible organisational decision to combine DeepMind and Google Brain into Google DeepMind. On paper, that should have concentrated the company’s best AI talent, reduced internal duplication and produced a clearer strategy. For a while, it appeared to work. Gemini 3 felt like a genuine recovery: the Pro model was capable, the pricing was aggressive compared with much of the competition, and Flash was useful when you needed a large amount of repetitive work completed cheaply. It still had the usual tendency to overthink simple instructions, but it was competitive. For a brief window, Google looked as though it might finally convert its enormous research advantage into a product advantage. That window did not remain open for very long.

AI Search launches by telling people to eat glue and jump off a bridge

Before Gemini became the centre of Google’s AI strategy, the company decided to put generated answers directly above ordinary search results. This was not an experimental chatbot hidden inside a research laboratory. It was Google Search, one of the most widely used and trusted information products in the world. People use Google to find medical guidance, financial information, instructions, news and answers during moments of genuine distress, so placing generated responses at the top of those results gave the system an enormous amount of implied authority.

The initial rollout went about as well as you would expect from modern Google. Its AI Overview told users to add non-toxic glue to pizza sauce to prevent the cheese from sliding off, apparently drawing the suggestion from an old joke posted on Reddit and presenting it as genuine cooking advice. It also produced the far more disturbing result in which a search from someone saying they felt depressed appeared to suggest jumping from the Golden Gate Bridge. Google said that some viral screenshots circulating at the time were fake and that many of the failures came from unusual or adversarial queries. That may have been true of some examples, but it did not erase the wider failure. Google had taken information from the open internet, removed much of its original context and presented the resulting text as a confident answer under Google’s own branding.

When traditional search gives someone a terrible result, the website is wrong. When an AI Overview synthesises terrible information and places it above every other result, Google is wrong. The company had decades of search experience, some of the best AI researchers alive and effectively unlimited testing resources, yet its introduction of generated answers became associated with glue pizza, eating rocks and apparent suicide advice. Google responded by adding safeguards, adjusting which sources the system relied upon and continuing with the rollout, but that response avoided the more important question of why a system carrying Google Search’s authority had been launched in that state at all. It became another example of the same pattern: announce the future, launch before the product is ready, endure public humiliation, install emergency restrictions and then move on without addressing the judgement that caused the failure.

Why is Google putting AI everywhere except where it makes sense?

Google owns the world’s dominant web browser, so Chrome should have been the most obvious place to demonstrate why Gemini could be useful to ordinary people. A capable browser assistant could understand the page someone is viewing, summarise long articles, compare several open tabs, explain technical language, extract dates, find buried information and assist with repetitive tasks. It is one of the clearest environments in which an average person could experience useful AI without first learning how to prompt a separate chatbot. Instead, Gemini’s deeper Chrome integration has been fragmented across regions, languages, accounts and subscription plans. Google owns the browser and has hundreds of millions of potential users already sitting inside it, but it still has not presented one coherent, universally available version of what Gemini in Chrome is supposed to be.

This is not exclusively a Google problem. Nearly every major technology company appears determined to put AI in the wrong places and then act confused when the public begins to hate it. Microsoft pushed Copilot branding across Windows without first establishing a compelling, reliable reason for ordinary users to interact with it. Rather than beginning with a few brilliant features that clearly improved the operating system, it attached AI to menus, buttons and products until Copilot became associated with clutter and forced integration. A considerable amount of anti-AI sentiment is not irrational resistance to technology. It is a direct reaction to companies adding mediocre AI to products people already liked, degrading the experience and then insisting that the degradation represents progress.

Google is creating the same problem inside Workspace. I use Gemini in Google Docs, so I know what the actual experience is like, and it is frequently a complete and utter fucking idiot. An AI writing assistant built directly into Docs should understand the document, retain its tone, make precise changes and help with the exact sentence or paragraph in front of the user. Instead, Gemini often behaves like a disconnected chatbot floating beside the document. It may technically have access to the content, but the experience rarely feels as though the AI genuinely understands the structure, purpose or voice of what is being written. That is not meaningful integration. It is a sidebar wearing an integration badge.

Gmail exposes the same strategic stupidity. Email is one of the most obvious places where AI could help ordinary people without forcing them to become chatbot enthusiasts. Long threads could be summarised automatically, decisions and requested actions could be extracted, dates could be added to a calendar, and routine replies could be drafted in the user’s normal style. Google owns the inbox, the draft, the contact history and much of the surrounding context, yet many of the useful Gemini features are restricted behind Google AI or eligible Workspace subscriptions. Someone who wants help writing a single email is not going to stop what they are doing, investigate Google’s subscription tiers and begin paying for another plan. They will open ChatGPT and type, “Write this email for me.” Google owns every part of the environment in which the email is being written and has still allowed another company’s chatbot to become the more obvious tool.

The mistake is treating every potentially useful AI interaction as another subscription opportunity rather than using several excellent features to demonstrate why integrated AI matters. Google should automatically summarise long threads, identify what requires a response, explain complicated documents, compare browser tabs and remove repetitive work. Those features should be reliable, visible and easy to understand. Users who want them should be able to enable them, while users who do not should be allowed to turn them off completely. AI should improve the existing product without obstructing it. Do not shove a chatbot into every corner of the interface, do not make the original application worse, and do not require a conversation with Gemini when a small automatic summary would solve the problem. Put AI where it actually makes sense.

Privacy that depends on which version of Gemini you happen to be using

Google’s privacy system is another example of something that may be technically explainable internally while remaining incoherent to the customer. With the ordinary personal Gemini app, activity settings determine whether conversations are retained in the user’s account and may contribute to product improvement, including through human review. Users can turn Gemini activity off, but chats may still be retained temporarily for service operation and safety, while conversations already selected for human review may be retained separately for considerably longer. That is not the same as saying every prompt is permanently used to train Gemini, and it would be inaccurate to claim that users have no control whatsoever. The problem is that understanding the actual rules requires navigating several settings pages, privacy notices and distinctions between products.

Those rules also change depending on whether Gemini is being used through the ordinary consumer application, inside Workspace, or through an organisation with enterprise protections. Google says that eligible Workspace content is not used to train its generative models, which is a significantly stronger privacy position. However, paying for an expensive consumer Google AI plan does not automatically make every Gemini interaction equivalent to using a managed enterprise Workspace account. A normal customer paying a substantial monthly fee could reasonably assume they have purchased the premium and more private version of the service, yet they still need to understand Google’s organisational divisions to determine what happens to their prompts.

If Google wants Gemini to access people’s inboxes, documents, browser tabs, files and personal context, this cannot remain a maze of product-specific qualifications. There should be a single obvious privacy control explaining in normal language whether conversations are saved, whether they contribute to model improvement and whether human review is allowed. Users should not need to understand the distinction between the Gemini app, Gemini in Workspace and enterprise Workspace protections before deciding whether to paste sensitive information into a prompt. Privacy should be a visible product feature, not a research project.

Antigravity and the classic limit-axe mistake

Around the same period, Google launched and heavily promoted Antigravity, its agentic development platform. It was presented as a glimpse of how software would be built with AI in the future, using multiple agents, richer views of ongoing work and tighter integration with Google’s models. Some parts of it were genuinely interesting. Pioneer Agent View was a good idea, and similar approaches have appeared in other coding tools. Google clearly understood that coding agents needed to become visible, manageable and collaborative rather than remaining chat boxes that occasionally edited files.

The problem was everything surrounding those ideas. The interface remained clunky, inconsistent and unfinished, feeling more like a collection of experiments than a product designed around a coherent development workflow. It contained impressive individual components, but the overall experience was frustrating enough that I would not recommend it to anyone I actually like. Google then made one of the most predictable product mistakes in the industry by giving free, Pro and Ultra users relatively generous access, allowing them to build habits around that access, and sharply cutting the limits once the operating costs became uncomfortable.

That is the worst possible way to introduce limits. A company should not provide a high-quality experience, allow people to depend on it and then suddenly pull the floor out from beneath them. It should establish realistic limits at launch, give free users a smaller number of genuinely useful runs and make the economics clear before people build their workflows around the platform. Instead, Google over-delivered, created dependency and then slammed the door. The damage was not simply that users received fewer tokens or agent runs. The real damage was trust. Once developers believe that allowances will be reduced after they adopt a platform, every generous new limit begins to look temporary. For a company asking people to build serious workflows around its products, that is catastrophic.

The Flash models that refuse to improve

Google repeatedly positions its Flash models as efficient workhorses for coding, agents and high-volume tasks. On paper, that should be one of the company’s biggest strengths. Google has enormous infrastructure, custom TPUs and decades of experience operating global systems, so it should be exceptionally capable of producing models that are fast, inexpensive and reliable. Yet the Flash line continues to encounter the same fundamental problems.

At Google I/O, Gemini 3.5 Flash was positioned as a major improvement for agentic work and coding. The marketing was strong, but across the workloads I tested, the actual experience was closer to a lobotomised Claude Haiku. It could generate plenty of tokens, but the problem was whether those tokens were useful. It repeatedly overthought straightforward instructions, wandered away from the requested task and produced responses that appeared busy without being precise. In coding workflows, it often felt less like an efficient agent and more like a model performing the appearance of one.

Google said it had listened to criticism, then released Gemini 3.6 Flash with the familiar promises of greater efficiency, stronger reliability and improved performance as a general workhorse. The core issues remained. It still overthinks simple tasks, struggles to follow instructions with the consistency expected from an agentic model and can spend more time describing what it plans to do than doing it. In real development work, these failures matter more than a benchmark improvement or a polished graph shown during a launch presentation.

Token efficiency also means very little if the model requires repeated correction. A model that consumes fewer tokens per response but needs three attempts to finish the job is not efficient. A more expensive model that understands the instruction and completes it correctly on the first attempt may be considerably cheaper in practice. This is where Google’s position becomes especially strange. It is not merely competing with OpenAI and Anthropic, but with increasingly capable Chinese open models that are often cheaper, more controllable and surprisingly effective inside a good coding harness. It is also competing with Meta, which had been relatively quiet at the frontier for a period and still managed to ship Muse Spark 1.1 as a credible model for coding, computer use and agentic work. Google, the company that created the transformer architecture, is somehow ending up on the wrong side of those comparisons.

Pricing that no longer makes sense

Early Gemini Flash pricing was one of the clearest reasons to tolerate the models’ quirks. They were cheap enough that unnecessary reasoning, occasional instruction failures and inconsistent output could be accepted as part of the trade-off. That advantage has weakened as parts of the Flash line have become more expensive and entered an awkward middle ground. The models are no longer always cheap enough to be the automatic workhorse, but they are not consistently capable enough to justify premium pricing.

A cheap and flawed model can still be extremely useful, just as an expensive and excellent model can be useful. The worst position is to be neither cheap enough to forgive nor good enough to trust, and that is increasingly where Gemini Flash sits. The relevant cost is not merely the advertised price per million tokens. It is the total cost of completing a successful job, including retries, corrections, unnecessary reasoning, failed tool calls and the developer time required to supervise the model. Once those costs are considered, a supposedly cheap model can become remarkably expensive. Google should understand this better than almost any company, yet it appears determined to optimise the pricing table while ignoring the actual experience behind it.

Gemma is the exception that proves the rule

None of this criticism applies neatly to Gemma. The Gemma models remain solid and, in some cases, excellent. They are useful locally, approachable for developers and generally easier to understand as products. The teams working on them appear to know what they are building and who they are building it for, which makes the contrast with the main Gemini line even more noticeable.

Gemma demonstrates that Google is still fully capable of shipping focused and technically impressive models with a clear purpose. The company has not lost its researchers or forgotten how to build good technology. The problem is that this talent becomes trapped inside a wider organisation that repeatedly changes names, limits, prices, priorities and product boundaries. Google does not lack intelligence or resources. It lacks consistency.

Google is competing with itself and losing

It would be wrong to claim that Google is not competing at all. Gemini is integrated throughout Search, Android and Workspace, Antigravity remains strategically important, and Google possesses enormous infrastructure alongside an almost unfair distribution advantage. The problem is that none of this feels coherent from the customer’s perspective. OpenAI has a recognisable identity around ChatGPT and its API, Anthropic has built one around Claude, coding and professional use, xAI has a clear identity whether people like it or not, and the strongest Chinese laboratories are competing aggressively on cost, openness and capability.

Google instead has Gemini in Chrome, Gemini in Search, Gemini in Workspace, Gemini Flash, Gemini Pro, Gemma, Antigravity, Google AI Pro, Google AI Ultra and a collection of overlapping features whose availability depends on the user’s country, account, subscription and product. It owns the browser but restricts the browser assistant. It owns Gmail but places some of the most obvious email features behind another subscription. It owns Docs but provides an assistant that often fails to understand the document properly. It wants access to deeply personal context while making its privacy system difficult to explain. It launches products with generous limits, cuts those limits after people adopt them and then appears surprised when developers stop trusting the platform.

That is the real fuck-up. It is not one bad model, one disastrous search result, one poor interface or one pricing change, but a repeated pattern in which Google invents something important, releases something promising, creates momentum and then undermines that momentum through confused execution. The company that helped create the architecture behind the modern AI industry continues to ship products that feel as though they are solving last year’s problems while the rest of the field moves forward. Google is not being defeated by a lack of resources. It is being defeated by Google.