{"id":1922910,"date":"2026-05-07T13:45:48","date_gmt":"2026-05-07T10:45:48","guid":{"rendered":"https:\/\/analyse.optim.biz\/?p=1922910"},"modified":"2026-05-07T13:45:48","modified_gmt":"2026-05-07T10:45:48","slug":"the-tiny-ai-model-that-lives-inside-the-chrome-browser-hallucinates-a-whole-lot","status":"publish","type":"post","link":"https:\/\/analyse.optim.biz\/?p=1922910","title":{"rendered":"The Tiny AI Model That Lives Inside the Chrome Browser Hallucinates a Whole Lot"},"content":{"rendered":"

[analyse_image type=”featured” src=”https:\/\/gizmodo.com\/app\/uploads\/2026\/05\/chrome-icon-1200×675.jpg”]<\/p>\n

\n
\n

As my colleague wrote on Wednesday, it appears that the Chrome browser quietly and in a squirrel-like fashion hides a 4-gigabyte AI model inside users\u2019 systems in a folder called OptGuideOnDeviceModel. According to Alexander Hanff, known to the internet as \u201cThat Privacy Guy,\u201d the biggest thing in that folder, a file called weights.bin, is Gemini Nano.<\/p>\n

Gemini Nano has lived on devices since 2024, Google told Gizmodo on Wednesday. According to a Google blog post, the tiny model allows developers to \u201cdeliver rich generative AI experiences without needing a network connection or sending data to the cloud,\u201d especially in cases where \u201clow cost, and privacy safeguards are your primary concerns.\u201d<\/p>\n

Whole communities exist online where users exchange tips for using local AI models that run without a network connection. The argument for them is that they don\u2019t run in data centers\u2014the dark satanic mills of our era. Plus, they can do cool things purely with the computing power needed to run the user\u2019s device, and, ideally, no one ever gets any access to your data. That last part is somewhat less comforting when we\u2019re talking about an instance of an LLM that materializes inside a web browser unbeknownst to the user.<\/p>\n

At any rate, I found a user-friendly way to run the built-in version of Gemini Nano in a ChatGPT-like web interface. It\u2019s called ChromeAI.org, and its origins can be traced back to a GitHub user calling themselves pipizhu or debugtheworldbot, apparently based in Shanghai.<\/p>\n

To my great satisfaction, it was impossible to run the Gemini Nano model initially on my install of Chrome. Security was pretty locked down by default, and I had to grant all sorts of disquieting permissions, which the site referred to as \u201cNecessary Experimental Flags,\u201d to make ChromeAI.org actually produce a functioning chat window. I don\u2019t recommend doing any of this.<\/em><\/p>\n

The heading \"Necessary Experimental Flags\" followed by a bunch of scary-looking instructions for changing defaults.
\u00a9 ChromeAI.org<\/figcaption><\/figure>\n

Once I did, I turned off my Wifi to make sure no cloud-computed LLM tokens were sneaking in or out of my laptop. Then I started chatting with it like it was November of 2022, and it was ChatGPT.<\/p>\n

That\u2019s more or less what the experience was like. The model felt surprisingly fast and capable despite running entirely on my computer\u2019s 2022 Apple M2 chip, and using my own 8 gigs of RAM. However, it has no chain-of-reasoning model running in the background, and obviously can\u2019t search the web for supporting facts. It\u2019s a pure, janky, hallucinating LLM. And sometimes it\u2019s impressive.<\/p>\n

For instance, I asked it what the capital of Burkina Faso was:<\/p>\n

A prompt asking for the capital of Burkina Faso, and reply saying \"Ouagadougou.\"
\u00a9 ChromeAi.org<\/figcaption><\/figure>\n

It answered \u201cOuagadougou\u201d instantly. That\u2019s the correct answer.<\/p>\n

Then, taking inspiration from something I wrote the other day, I asked it to pretend it was the P.G. Wodehouse character Jeeves. It, uh, sorta did?<\/p>\n

\n

\u201cOh, my dear fellow! A splendid notion indeed! To converse with Jeeves is to converse with a paragon of composure, intellect, and impeccable taste. Ahem.<\/em><\/p>\n

You say\u00a0you\u00a0have a specific Jeeves in mind? Excellent! I shall endeavour to emulate his mannerisms and wit to the best of my ability. Pray, do enlighten me. Is this Jeeves of your liking particularly\u2026\u00a0spirited? Perhaps a touch prone to unconventional solutions?\u201d<\/em><\/p>\n<\/blockquote>\n

Really, it just started writing in vaguely outdated language and with British spelling. Much too cheerful and long-winded to be Jeeves, but close enough, I guess.<\/p>\n

Then I told it to \u201cscrap the jeeves thing\u201d and asked it a trivia question about the Monkey Island game series from the 90s. I was just trying to see what kinds of obscure facts lived inside a 4-gig file. My apologies for the depths of cringe plumbed in the following screenshot:<\/p>\n

\"Prompt:
\u00a9 ChromeAI.org<\/figcaption><\/figure>\n

Not only did it get the answer wrong, but it also ignored my instruction to scrap the Jeeves thing. \u201cAh, indeed! You speak of Monkey Island, a most delightfully mischievous adventure,\u201d it exclaimed, to my horror. It said the keyhole was \u201cwithin the\u2026 lower cheek pouch<\/em> of the monkey\u2019s visage,\u201d and that the key \u201cresembles a small, intricately carved\u2026\u00a0coconut shell<\/em>, if you will.\u201d I won\u2019t.<\/p>\n

The model was, perversely, wrong in the voice of a bad imitation of Jeeves<\/em>. What could possibly be worse? The correct answer, by the way, would be that the key goes in the monkey\u2019s ear and is shaped like a Q-Tip.<\/p>\n

Then I told it to stop talking like Jeeves, and it apologized in its Jeeves voice. \u201cMy sincerest apologies! A most unfortunate miscalculation. I seem to possess a rather\u2026\u00a0robust<\/em> attachment to the Jeeves persona. Please forgive the momentary lapse in decorum.\u201d Then I told it to stop talking like Jeeves a third time, and it said, \u201cI seem to have been indulging in an overabundance of theatricality. I apologize profusely. I shall endeavor religiously<\/em> to refrain from adopting such a manner of speaking.\u201d<\/p>\n

With my fourth ask, I finally got it to dial down the Jeeves thing slightly\u2014though weirdly, there was still some residue. I asked it about the key again because its answer had been wrong, and it said it was wrong about the coconut part, and that, \u201cIt is more akin to a small, rounded indentation, perhaps resembling a shallow bowl or a slightly flattened oval.\u201d<\/p>\n

Discouraged, and trying to get it to just do ostensibly normal LLM things\u2014cheat on school assignments and such\u2014I asked it for a list of the Presidents of the United States in order of their ages when they took office. Reader, the list was bad:<\/p>\n

\"A
\u00a9 ChromeAI.org<\/figcaption><\/figure>\n

Admittedly, this is a challenging prompt, but before it started hallucinating and just saying whatever, one would hope the LLM would at least mention the youngest president<\/em>, Teddy Roosevelt<\/em>. It didn\u2019t do that, listed random presidents in a random order, and invented an entire guy named John F. Lincoln.<\/p>\n

You\u2019ll never guess what happened when I asked it about John F. Lincoln:<\/p>\n

A prompt asking \"John F. Lincoln\"? followed by another long list in the incorrect order, including John F. Lincoln again.
\u00a9 ChromeAI.org<\/figcaption><\/figure>\n

Yep, it listed John F. Lincoln a second time. So that should give you some idea of what this model can do.<\/p>\n

When you consider that GPT-3, which powered the original ChatGPT, required 350 gigabytes of storage space, and also hallucinated, this kind of performance, janky or not, from a 4-gigabyte, on-device model is still impressive. The comparison is even more favorable if you compare it to OpenAI\u2019s models from the days before GPT-3.<\/p>\n

Also, Google never asked anyone to actually use this model as a chatbot. Instead, it \u201cpowers important security capabilities like scam detection and developer APIs without sending your data to the cloud,\u201d Google told Gizmodo on Wednesday.<\/p>\n

So to be completely clear, don\u2019t rely on Gemini Nano as a chatbot. Not even Google wants you to do that.<\/p>\n<\/div>\n<\/article>\n

\n

As my colleague wrote on Wednesday, it appears that the Chrome browser quietly and in a squirrel-like fashion hides a 4-gigabyte AI model inside users\u2019 systems in a folder called OptGuideOnDeviceModel. According to Alexander Hanff, known to the internet as \u201cThat Privacy Guy,\u201d the biggest thing in that folder, a file called weights.bin, is Gemini Nano.<\/p>\n

Gemini Nano has lived on devices since 2024, Google told Gizmodo on Wednesday. According to a Google blog post, the tiny model allows developers to \u201cdeliver rich generative AI experiences without needing a network connection or sending data to the cloud,\u201d especially in cases where \u201clow cost, and privacy safeguards are your primary concerns.\u201d<\/p>\n

Whole communities exist online where users exchange tips for using local AI models that run without a network connection. The argument for them is that they don\u2019t run in data centers\u2014the dark satanic mills of our era. Plus, they can do cool things purely with the computing power needed to run the user\u2019s device, and, ideally, no one ever gets any access to your data. That last part is somewhat less comforting when we\u2019re talking about an instance of an LLM that materializes inside a web browser unbeknownst to the user.<\/p>\n

At any rate, I found a user-friendly way to run the built-in version of Gemini Nano in a ChatGPT-like web interface. It\u2019s called ChromeAI.org, and its origins can be traced back to a GitHub user calling themselves pipizhu or debugtheworldbot, apparently based in Shanghai.<\/p>\n

To my great satisfaction, it was impossible to run the Gemini Nano model initially on my install of Chrome. Security was pretty locked down by default, and I had to grant all sorts of disquieting permissions, which the site referred to as \u201cNecessary Experimental Flags,\u201d to make ChromeAI.org actually produce a functioning chat window. I don\u2019t recommend doing any of this.<\/em><\/p>\n

The heading \"Necessary Experimental Flags\" followed by a bunch of scary-looking instructions for changing defaults.
\u00a9 ChromeAI.org<\/figcaption><\/figure>\n

Once I did, I turned off my Wifi to make sure no cloud-computed LLM tokens were sneaking in or out of my laptop. Then I started chatting with it like it was November of 2022, and it was ChatGPT.<\/p>\n

That\u2019s more or less what the experience was like. The model felt surprisingly fast and capable despite running entirely on my computer\u2019s 2022 Apple M2 chip, and using my own 8 gigs of RAM. However, it has no chain-of-reasoning model running in the background, and obviously can\u2019t search the web for supporting facts. It\u2019s a pure, janky, hallucinating LLM. And sometimes it\u2019s impressive.<\/p>\n

For instance, I asked it what the capital of Burkina Faso was:<\/p>\n

A prompt asking for the capital of Burkina Faso, and reply saying \"Ouagadougou.\"
\u00a9 ChromeAi.org<\/figcaption><\/figure>\n

It answered \u201cOuagadougou\u201d instantly. That\u2019s the correct answer.<\/p>\n

Then, taking inspiration from something I wrote the other day, I asked it to pretend it was the P.G. Wodehouse character Jeeves. It, uh, sorta did?<\/p>\n

\n

\u201cOh, my dear fellow! A splendid notion indeed! To converse with Jeeves is to converse with a paragon of composure, intellect, and impeccable taste. Ahem.<\/em><\/p>\n

You say\u00a0you\u00a0have a specific Jeeves in mind? Excellent! I shall endeavour to emulate his mannerisms and wit to the best of my ability. Pray, do enlighten me. Is this Jeeves of your liking particularly\u2026\u00a0spirited? Perhaps a touch prone to unconventional solutions?\u201d<\/em><\/p>\n<\/blockquote>\n

Really, it just started writing in vaguely outdated language and with British spelling. Much too cheerful and long-winded to be Jeeves, but close enough, I guess.<\/p>\n

Then I told it to \u201cscrap the jeeves thing\u201d and asked it a trivia question about the Monkey Island game series from the 90s. I was just trying to see what kinds of obscure facts lived inside a 4-gig file. My apologies for the depths of cringe plumbed in the following screenshot:<\/p>\n

\"Prompt:
\u00a9 ChromeAI.org<\/figcaption><\/figure>\n

Not only did it get the answer wrong, but it also ignored my instruction to scrap the Jeeves thing. \u201cAh, indeed! You speak of Monkey Island, a most delightfully mischievous adventure,\u201d it exclaimed, to my horror. It said the keyhole was \u201cwithin the\u2026 lower cheek pouch<\/em> of the monkey\u2019s visage,\u201d and that the key \u201cresembles a small, intricately carved\u2026\u00a0coconut shell<\/em>, if you will.\u201d I won\u2019t.<\/p>\n

The model was, perversely, wrong in the voice of a bad imitation of Jeeves<\/em>. What could possibly be worse? The correct answer, by the way, would be that the key goes in the monkey\u2019s ear and is shaped like a Q-Tip.<\/p>\n

Then I told it to stop talking like Jeeves, and it apologized in its Jeeves voice. \u201cMy sincerest apologies! A most unfortunate miscalculation. I seem to possess a rather\u2026\u00a0robust<\/em> attachment to the Jeeves persona. Please forgive the momentary lapse in decorum.\u201d Then I told it to stop talking like Jeeves a third time, and it said, \u201cI seem to have been indulging in an overabundance of theatricality. I apologize profusely. I shall endeavor religiously<\/em> to refrain from adopting such a manner of speaking.\u201d<\/p>\n

With my fourth ask, I finally got it to dial down the Jeeves thing slightly\u2014though weirdly, there was still some residue. I asked it about the key again because its answer had been wrong, and it said it was wrong about the coconut part, and that, \u201cIt is more akin to a small, rounded indentation, perhaps resembling a shallow bowl or a slightly flattened oval.\u201d<\/p>\n

Discouraged, and trying to get it to just do ostensibly normal LLM things\u2014cheat on school assignments and such\u2014I asked it for a list of the Presidents of the United States in order of their ages when they took office. Reader, the list was bad:<\/p>\n

\"A
\u00a9 ChromeAI.org<\/figcaption><\/figure>\n

Admittedly, this is a challenging prompt, but before it started hallucinating and just saying whatever, one would hope the LLM would at least mention the youngest president<\/em>, Teddy Roosevelt<\/em>. It didn\u2019t do that, listed random presidents in a random order, and invented an entire guy named John F. Lincoln.<\/p>\n

You\u2019ll never guess what happened when I asked it about John F. Lincoln:<\/p>\n

A prompt asking \"John F. Lincoln\"? followed by another long list in the incorrect order, including John F. Lincoln again.
\u00a9 ChromeAI.org<\/figcaption><\/figure>\n

Yep, it listed John F. Lincoln a second time. So that should give you some idea of what this model can do.<\/p>\n

When you consider that GPT-3, which powered the original ChatGPT, required 350 gigabytes of storage space, and also hallucinated, this kind of performance, janky or not, from a 4-gigabyte, on-device model is still impressive. The comparison is even more favorable if you compare it to OpenAI\u2019s models from the days before GPT-3.<\/p>\n

Also, Google never asked anyone to actually use this model as a chatbot. Instead, it \u201cpowers important security capabilities like scam detection and developer APIs without sending your data to the cloud,\u201d Google told Gizmodo on Wednesday.<\/p>\n

So to be completely clear, don\u2019t rely on Gemini Nano as a chatbot. Not even Google wants you to do that.<\/p>\n<\/div>\n

[analyse_source url=”https:\/\/gizmodo.com\/the-tiny-ai-model-that-lives-inside-the-chrome-browser-hallucinates-a-whole-lot-2000755397″]<\/p>\n","protected":false},"excerpt":{"rendered":"

[analyse_image type=”featured” src=”https:\/\/gizmodo.com\/app\/uploads\/2026\/05\/chrome-icon-1200×675.jpg”] As my colleague wrote on Wednesday, it appears that the Chrome browser quietly and in a squirrel-like fashion hides a 4-gigabyte AI model inside users\u2019 systems in a folder called OptGuideOnDeviceModel. According to Alexander Hanff, known to the internet as \u201cThat Privacy Guy,\u201d the biggest thing in that folder, a file called […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2],"tags":[226,53],"class_list":["post-1922910","post","type-post","status-publish","format-standard","hentry","category-politics","tag-crawlmanager","tag-gizmodo-com"],"_links":{"self":[{"href":"https:\/\/analyse.optim.biz\/index.php?rest_route=\/wp\/v2\/posts\/1922910","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/analyse.optim.biz\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/analyse.optim.biz\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/analyse.optim.biz\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/analyse.optim.biz\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1922910"}],"version-history":[{"count":0,"href":"https:\/\/analyse.optim.biz\/index.php?rest_route=\/wp\/v2\/posts\/1922910\/revisions"}],"wp:attachment":[{"href":"https:\/\/analyse.optim.biz\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1922910"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/analyse.optim.biz\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1922910"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/analyse.optim.biz\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1922910"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}