For the Next Month: Going All-in on The Browser Company's Dia-AI
& we will see where we get, and whether we need… ROADS! From searchbox to sidekick: is the humble web browser the secret to mastering the potential of MAMLMs as natural-language interfaces and as...
& we will see where we get, and whether we need… ROADS! From searchbox to sidekick: is the humble web browser the secret to mastering the potential of MAMLMs as natural-language interfaces and as superbig-data, superhigh-dimension, superflexible-function classifiers, rather than being mastered by it? Perhaps I can tune Dia Browser so that it truly remaps the web for my cognitive benefit…
I have decided to embark on a month-long experiment going all-in using Dia Browser <https://www.diabrowser.com/>.
I have decided to explore if I can actually use GPT LLM MAMLM ChatBots—General-Purpose Transformer Large-Language Model Modern Advanced Machine-Learning Model ChatBots—to manage my current very bad case of information overload.
I also hope it will help me critically assessing the real limits and potential of MAMLMs.
Specifically, I hope to gain some insight into how I might be wrong in my current judgment that they can only be of substantial yet strangely limited usefulness—that their simulacrum of intelligence is in fact a place where Sokrates’s observation that no matter how much a statue looks like the person it cannot speak is in fact correct. Think of them as, say, doing for prose roughly what spreadsheets did for accounting—important, mind-blowing, but not Chicxulub.
Back up:
General-Purpose Transformer Large-Language Model Modern Advanced Machine-Learning Models—GPT LLM MAMLMs—are primarily:
Very big-data, very high-dimension, very flexible-function classification analyses.
Natural-language interfaces to structured and unstructured databases.
And then the rest of the technological and societal knock-on from their coming:
Platform oligopolists spending money and resources at a scale relative to the then-size of the economy that has never before seen in any General-Purpose Technology build-out…
Why? Not to make or to try to make more money, but to buy insurance to guard against Christensenian disruption of their current platform-oligopoly profit flows…
Not FILTH—Failed in London? Try Hong Kong!—but FICTA—Failed in Crypto? Try AI!: The post-crypto boom way of separating gullible VC from their money for the benefit of engineers, speculators, utopians, and grifters…
Auto-complete for everything on super-steroids…
Highly verbal electronic-software therapists, sounding boards, coaches, pets, and… THIS IS A SAFE-FOR-WORK WEBLOG!…
I have written about this a bunch. Most recently: for the Milken Institute Review here: <https://www.milkenreview.org/articles/behind-the-hype?IssueID=58>
I recommend that piece highly. But I had more to say that would not fit into that piece! And that got left ln the cutting-room floor!
Specifically, about GPT LLM MAMLM ChatBots:
GPT LLM MAMLM ChatBots merge the first two, the primary aspects of MAMLMs, to create a new summarization cultural technology for the internet, plus whatever else is in their training data.
I think of GPT LLM MAMLM ChatBots doing for prose roughly what spreadsheets did for accounting—important and mind-blowing, and a destroyer of many not-that-interesting jobs. But not world-shaking.
That is what I think.
And I think I have good reasons for thinking so.
But I might be wrong.
I think that because the core of a GPT LLM ChatBot, after all, is a huge neural network that is trained to emulate the language-response patterns found in its training data. In practice, that means that its telos is to become a closer and closer approximation of the typical Internet s***poster.
Yes, in addition to neural-network initial training, there is post-training RLHG, RAG, DPO, PEFT, LoRA, PPO, SFT, RM, IFT, and I am sure there will be at least five more such acronyms in common use in a year. But these are all merely (merely!) rough-and-ready ways to try to get the typical internet s***poster that is the ideal GPT LLM ChatBot to behave: to nudge it into a state in which the training data it sees around its attentive core is dense, and in which that training data was constructed by smart, informed people trying to communicate true and important stuff. Actively trying to strip the thing down so that it does not wander off into internet woo-woo land—that is an active area of “research”.
Witness Andrej Karpathy here:
Andrej Karpathy: <https://x.com/karpathy/status/1938626382248149433/>: ‘The race for LLM "cognitive core" - a few billion param model that maximally sacrifices encyclopedic knowledge for capability. It lives always-on and by default on every computer as the kernel of LLM personal computing. Its features are slowly crystalizing…. atively multimodal text/vision/audio at both input and output. - Matryoshka-style architecture…. Reasoning, also with a dial.… On-device finetuning LoRA slots for test-time training, personalization and customization…. It doesn't know that William the Conqueror's reign ended in September 9 1087, but it vaguely recognizes the name and can look up the date. It can't recite the SHA-256 of empty string as e3b0c442..., but it can calculate it quickly…. Super low interaction latency… private access to data and state, offline continuity, sovereignty ("not your weights not your brain")…. Do people *feel* how much work there is still to do. Like wow.
And here:
Andrej Karpathy: <https://x.com/karpathy/status/1937902205765607626>: ‘+1 for "context engineering" over "prompt engineering"…. In every industrial-strength LLM app… the delicate art and science of filling the context window with just the right information…. Science because doing this right involves task descriptions and explanations, few shot examples, RAG, related (possibly multimodal) data, tools, state and history, compacting.… Art because of the guiding intuition around LLM psychology of people spirits…. Vreak up problems just right into control flows - pack the context windows just right - dispatch calls to LLMs of the right kind and capability - handle generation-verification UIUX flows - a lot more…. The term "ChatGPT wrapper" is tired and really, really wrong…
Thus I tend to dissent from things like Torsten Slok’s claim that “AI” is going to deliver faster measured real GDP growth here in the U.S. over the next five years:
And I have explained why I see the GPT LLM category of MAMLMs as limited to roughly spreadsheet—far below microprocessor—levels of impact:
I see them as such because, at bottom, I see them as what I was taught in third-grade math to call “function machines”: essentially “autocomplete on steroids”. (It is still the case that the best comprehensible introduction to this (for me at least) that I have been pointed to is still Alessandrini & al. (2023).) But. t
The results are very, very impressive indeed across a number of dimensions. And every time I reach this point, I am again unable to avoid quoting Shalizi (2023, 2025):
Cosma Shalizi: ‘Attention’, ‘Transformers’, in Neural Network ‘Large Language Models’ <http://bactra.org/notebooks/nn-attention-and-transformers.html>: ‘"It's Just Kernel Smoothing" vs. "You Can Do That with Just Kernel Smoothing!?!": [That] takes nothing away from the incredibly impressive engineering accomplishment of making the blessed thing work…. nobody [before] achieved anything like the[se] feats…. [We] put effort into understanding… precisely because the results are impressive!…
The neural network architecture here is doing some sort of complicated implicit smoothing across contexts… [that] has evolved (under the selection pressures of benchmark data sets and beating the previous state-of-the-art) to work well for text as currently found online.… Markov models for language are really old…. Nobody, so far as I know, has achieved results anywhere close to what contemporary LLMs can do. This is impressive...
So I am left with a worrisome, nagging feeling: these things are unreasonably effective.
Why are large language models (LLMs) so astonishingly effective?
I guess: primarily because of scale. Trained on vast datasets and operating in thousands of dimensions, they capture the statistical patterns of human language. When you request a summary or script, the output is not true understanding, but a “cultural average”—how people have written such things before. The high dimensionality of these models’ internal representations is important too. Doing math like “king + woman - man ≈ queen” is not just a clever parlor trick, but evidence of deeper structural regularities and even meaning encoded in the vectors and matrices. There is thus an emergent flexibility and richness come from the neural network’s ability to classify and interpolate across a vast, multidimensional space.
What does this mean in practice? LLMs excel at compressing, interpolating, and representing the latent structures of human language and knowledge. They are, as the phrase goes, “autocomplete on steroids.” Their outputs are plausible, coherent, and often insightful, even though they lack genuine understanding or grounding in reality.
In my view, LLMs’ “unreasonable effectiveness” is likely to have arisen from five main factors.
First, their scale—massive data and computation—enables them to learn what hand-crafted rules cannot.
Second, their statistical machinery allows them to interpolate between immense numbers of examples, producing plausible continuations for prompts that bear only a family resemblance to things they have seen.
Third, their training uncovers latent structures that make language compressible and predictable, echoing Plato’s dream of “carving
naturelanguage at its joints.”Fourth, their high-dimensional embeddings allow for analogical reasoning that mimics superficially much the way humans think.
Fifth—and perhaps most important—a huge amount of real human thought is preserved in crystalized language pattern-correlations.
Stil, with all of that: A great deal of time, energy, and effort has been spent training my 50W-drawing wetware that fits inside the small breadbox-sized bone enclosure at the top of my spine. That had better be better at true cognition—at searching out, absorbing, processing, and then signaling and conveying information. It is, after all, a full Turing-Class entity. If a blank-slate neural network orders-of-magnitude less complex trained to emulate a typical internet s***poster can steam-hammer-to-John-Henry me, that time, energy and effort has been largely wasted. That seems very unlikely. No. If you are actually going to build a Turing-Class sEntity, attention, scale, and a blank-slate neural network to train are highy unlikely to be all that you need.
Thus think of them as the next step in the long tradition of tools that let us offload rote, repetitive, or formulaic parts of knowledge work, freeing up our attention for higher-order thinking. It is the next step in a process that started back when the first task of a scribe in writing a document was mixing the clay for the cuneiform to the proper consistency and smoothness, and the sudden ability to import papyrus to Sumer from Tawy was a true godsend.
And yet, suppose I am right that GPT LLM ChatBots are merely neural-network Markov-chain without AGI emergence Sub-Turing mimics of human minds.
They are, still, very powerful cultural technologies for search and summarization above all. And for MOAR: also for link-following, code-snippet generating, ritual, milestone-marking, translation, genre-riffing, outlining, brainstorming, and more.
These models are not reasoning agents. They cannot verify facts and are prone to hallucinations—confidently generating plausible-sounding nonsense. But they do have strengths. They are calculators for language, not philosophers. They are tools for offloading cognitive drudgery, not engines of new insight. This is a real and a big productivity gain, especially for those who already know what “good” looks like and simply need a first draft or a nudge in the right direction. They are not magical.
But, still: They are remarkable. And useful.
GPT LLMs should be considered remixers and redistributors of existing culture, offering highly powerful affordances for access, recombination, and dissemination. There are dangers: the sheer persuasive power of a natural-language interface to humans tuned to human-sounding language primed to over-attribute intentionality. But if we guard against those, all of a sudden we now have a better way of funneling the firehose of information into a form in which we can make sense of it. We can distill the vast, chaotic sprawl of the territory human textual production into a manageable, if inevitably lossy, map.
And while the map is not the territory, some maps can be very good and useful indeed.
Thus these things are immensely useful for search and research. I do not know to what degree to blame Google executives’ pursuit of advertising profits at all cost, and to what degree to blame SEO search-arbitrage attention jackals for the degeneration of Google search. But even in its present, something like ChatGPT is a better search engine than Google search as long as you tell it to act like a high-quality search engine. Witness Joe Weisenthal:
Joe Weisenthal: ChatGPT Is Becoming My Default Search Engine <https://www.bloomberg.com/news/newsletters/2025-07-17/chatgpt-is-becoming-my-default-search-engine>: ‘The same type of query on Google isn’t nearly as useful…. The actual links that Google turns up (the old, core search business) is not nearly as helpful…. To be fair… Google is better… if I just want to find a Wikipedia page… or if I’m looking for a very specific phrase…. But for a range of queries, o3 from ChatGPT increasingly feels like a strictly better experience…
The act of using a GPT LLM to query the internet should not be thought of as an act of delegating thought, but rather of curating and steering a process of retrieval from the memory of the Anthology Super-Intelligence that is Humanity’s Collective Mind. collective memory retrieval; the user’s judgment remains indispensable. GPT LLMs promise to democratize access to cultural and intellectual resources, enabling non-experts to interact with and synthesize information in ways previously reserved for the highly literate or technically skilled. They also promise to greatly amplify the cognitive information-surveying and information-acquisition capabilities of those skilled in their use.
They are, I think, going to provide us with much better whips and chairs for trying to tame informational overload than we had even five short years ago.
Thus whether or not I am wrong about the limits of these ChatBots, natural-language interfaces to structured and unstructured datastores promise to be very useful indeed to somebody like me who is both a huge user of datastores and a very sophisticated user of natural language.
So how can I quickly and easily gain the maximum amount of facility with these tools so I can use them for good, and not either let the potential benefit I might draw from them wither on the vine, or, worse, find somebody else using them to hack my brain for my detriment?
(And also if I am wrong, learn quickly that I need to start practicing for when I will have to, in five years, bend the knee to our AI-Overlords.)
These issues are now even more tremendously important because we do now have true information overload.
This problem has been building for quite a while. If one is a white-collar worker—if one is a front-end node to and in the Anthology Super-Intelligence of Humanity’s Collective Mind—one’s tasks, whether it is that of trying to enrich or emprestige your employers so they pay you, or enrich your life—fall into three groups:
The business of thinking through an issue once you have gathered your sources of information.
The issue of then getting your conclusions disseminated in a form in which they are neither evanescent nor unpersuasive.
The issue of finding and assembling the information that you need to actually do your thinking.
Think of these as the signal, the process, and the funnel. Someone who wants to succeed and be useful to society in my business needs to be able to do all three, and to be able to do all three very well. I want very much to be able to do a better job in making sure that I read more of what I should be reading and from people who are not in the circle of usual suspects
The problem is the funnel. Dealing with the funnel is becoming an increasingly large part of our jobs. And here is where the hope that our modern information technologies understood as cultural technologies of summarization and classification may of great help.
I do not understand how these things work (but does anybody?). I do not understand how to use them well. There is too much to read. There is too much to experiment with. And yet, as Cosma Shalizi says, we “clearly… need to wrap… [our] head[s] around it, before… [we soon] become technically obsolete…”
And so I am going to try to spend the next month going all-in on The Browser Company’s Dia Browser <https://www.diabrowser.com/>. What is Dia Browser? One of legion coming down the pike:
M. G. Seigler: Begun, the AI Browser Wars Have <https://spyglass.org/ai-browser-wars/?ref=the-spyglass-column-newsletter: ‘Dia ushers in a new day for the web browser.… You could basically just use Dia as a slightly cleaner-looking version of Chrome… and never touch the AI elements…. [But] if Dia is going to succeed, it will be because of AI. As such, they're now in a foot race…. Google… every other browser… dabbling in some level of AI… Perplexity and… OpenAI…. Anyone who uses Dia should see why all of these players will build their own browsers. Extensions aren't going to be good enough here, you have to go deeper to fully control the experience and have full vision into what's going on in that browser…
Dia Browser is, with a “@” denoting a link to another browser tab:
Chat with your tabs… Hey Dia…
Write with your tabs… Make this @JuneNewsletter sound on-brand @BrandVoice…
Learn with your tabs… Explain this to me like I’m five @Singularities…
Plan with your tabs… Compare @TuscanVilla @IlCasone in a table with price, proximity, to best coffee…
Shop with your tabs… Convince me not to buy this @VintageFendiBaguette… <https://www.diabrowser.com/>
The idea is that with every webpage you visit you can then immediately, seemlessly ask a question related to it, also bringing in whatever else you think is potentially relevant to the answer you get. The hope is to make the interaction-feedback loop as tight as possible, and then see what develops.
Dia is The Browser Company’s second project. The first was their power-user browser Arc: <https://arc.net/> “The Chrome Replacement You’ve Been Waiting For!” The Browser Company’s ambition with Arc Browser was never just technical—it was anthropological. Their goal was to make the browser feel like a home on the internet, personal and customizable, letting users shape the technology to fit their lives to extend and adapt their browser into a digital home.
Dia Browser aims at this with a twist: empowering users—even those with no coding experience—to build their own tools and workflows by what is now called “vibe coding”. The bet is that one of the true promises of MAMLMs is in providing natural-language in-the-moment tool-building for simple bespoke software to help us navigate the messy, context-rich decisions of real life—like planning a trip, balancing preferences, or making sense of ambiguity.
Traditionally, browsers have served as “funnels for intent”—the search box where users express what they want—and as platforms for running webapps. But why should the browser omnibox link only to legacy search engines? Why not route intent to other digital experiences, webapps, including those powered by AI? Perhaps most value now lies not in the browser itself, but in what can be built atop it. The goal is a personal assistant—an “intelligence layer”—that sits across all devices, learns from user context, and helps manage tasks, not just tabs. In this new landscape, context, memory, and personalization may be the true differentiators and wisdom amplifiers.
The Browser Company found that users of the alpha version quickly found their way to writing their own simple prompts for the ChatBot in the right pane of the browser window to act on the webpage in the left to automate or streamline their daily tasks: new tools, workflows, and automations, mini-agents and micro-applications that users could build for themselves, or frictionlessly adopt and then modify.
Perhaps the browser is about to undergo a radical transformation in this age of MAMLMs. Perhaps a good way for it to take advantage of natural-language interface and stupendous classification capatilities is for it to stand in the middle, intercepting and interpreting my intent, turning the searchbox into a launchpad for personalized, context-aware assistance and assistants. Perhaps context, memory, and customization become key. The Browser Company thinks its Dia Browser could soon become my most valuable ally in dealing with information overload.
Is The Browser Company right?
I think I need to experiment enough to see.
References:
Alessandrini, Giulio, Brad Klee, & Stephen Wolfram. 2023. "What Is ChatGPT Doing ... and Why Does It Work?". <https://writings.stephenwolfram.com/2023/02/what-is-chatgpt-doing-and-why-does-it-work/>.
Browser Company of New York. 2025. “Meet Dia – the AI Browser Where You Can Chat with Your Tabs.” Dia Browser. https://www.diabrowser.com/.
Chiang, Ted. 2023. “ChatGPT Is a Blurry JPEG of the Web.” The New Yorker, February 9. https://www.newyorker.com/tech/annals-of-technology/chatgpt-is-a-blurry-jpeg-of-the-web.
DeLong, J. Bradford. 2025. “Behind the Hype: What ‘AI’ Is & Isn’t.” Milken Institute Review, Q3 2025. https://www.milkenreview.org/articles/behind-the-hype?IssueID=58.
DeLong, J. Bradford. 2025. “MAMLM as a General Purpose Technology: The Ghost in the GDP Machine.” Grasping Reality. June 24. https://braddelong.substack.com/p/mamlm-as-a-general-purpose-technology.
DeLong, J. Bradford. 2025. “Spreadsheet, Not Skynet: Microdoses, Not Microprocessors.” Grasping Reality. June 24. https://braddelong.substack.com/p/spreadsheet-not-skynet-microdoses.
Heath, Alex. 2025. “Perplexity’s CEO on why the browser is AI’s killer app”. The Verge. July 17. https://www.theverge.com/decoder-podcast-with-nilay-patel/708256/perplexity-ceo-decoder-aravind-srinivas-comet-browser-ai-search.
Holman, Mat. 2025 “AI means the end of internet search as we’ve known it”. MIT Technology Review. January 6. https://www.technologyreview.com/2025/01/06/1108679/ai-generative-search-internet-breakthroughs/.
Honan, Mat. 2025. “AI means the end of internet search as we’ve known it” MIT Technology Review, January 6. https://www.technologyreview.com/2025/01/06/1108679/ai-generative-search-internet-breakthroughs/.
Karpathy, Andrej. 2025. “+1 for ‘context engineering’ over ‘prompt engineering’.” X. June 25. https://x.com/karpathy/status/1937902205765607626.
Karpathy, Andrej. 2025. “The race for LLM ‘cognitive core’.” X. June 27. https://x.com/karpathy/status/1938626382248149433/.
Purohit, Rhea. 2025. “Inside The Browser Company: Why They Killed Arc To Build Dia.” Every. July 2. https://every.to/podcast/inside-the-browser-company-why-they-killed-arc-to-build-dia
OpenAI. 2025. “Exclusive: OpenAI to release web browser in challenge to Google Chrome.” Reuters, July 9. https://www.reuters.com/business/media-telecom/openai-release-web-browser-challenge-google-chrome-2025-07-09/
Perplexity AI. 2025. “Introducing Comet: Browse at the Speed of Thought.” Perplexity Blog. https://www.perplexity.ai/hub/blog/introducing-comet.
Reid, Elizabeth. 2025. “AI in Search: Going beyond information to intelligence.” Google Blog, May 20. https://blog.google/products/search/google-search-ai-mode-update/
Seigler, M. G. 2025. “The AI Browser Wars: Why Every Tech Giant Wants In.” Spyglass, July 16. https://spyglass.org/ai-browser-wars/?ref=the-spyglass-column-newsletter.
Shalizi, Cosma. 2025. “‘Attention’, ‘Transformers’, in Neural Network ‘Large Language Models’.” Bactra Notebooks. July 10. http://bactra.org/notebooks/nn-attention-and-transformers.html.
Shalizi, Cosma. 2025. “On Feral Library Card Catalogs, or, Aware of All Internet Traditions.” Programmable Mutter. April 17. https://www.programmablemutter.com/p/on-feral-library-card-catalogs-or.
Weisenthal, Joe. 2025. “ChatGPT Is Becoming My Default Search Engine.” Bloomberg July 17. https://www.bloomberg.com/news/newsletters/2025-07-17/chatgpt-is-becoming-my-default-search-engine.
I'm an old guy and can remember many things that err going to change everything, but didn't. Automation was going to end humans building cars. Mortgage backed securits were totally secure. The new freeways in LA were going to end traffic congestion. Allen Kamp
I look forward to your report.
Include any thoughts on how you think enshittification will likely proceed - e.g., subsequently charging for useful features such as cloud storage of documents, or particular AI-assisted features (Grammarly is doing that). This is not just an ongoing issue, but appears to be expanding in use. If the service shuts down, will there be any way to retrieve documents in the cloud? Can the service work offline or must it be used only online?