This is Brad DeLong's Grasping Reality—my attempt to make myself, and all of you out there in SubStackLand, smarter by writing where I have Value Above Replacement and shutting up where I do not… An Interim Report on My Experience with Dia Browser Use, & MAMLMs More GenerallyContexts, copilots, and chatbots: A mid-course report on the AI-augmented web…
Contexts, copilots, and chatbots: A mid-course report on the AI-augmented web…What happens as your browser evolves to become less of a passive portal? I am now a third of the way through my month of attempting to go all-in on Dia Browser and its WebChatBot copilot features. I still find myself very intrigued by the idea of having a natural-language interface to a copilot-assistant—like talking to yourself in the mirror, but the self you are talking to can use google, and do a lot of semi-automatic text processing for you as well. I still do not have any no strong conclusions. But I can say, the arrival of “AI” webbrowser-copilots like Dia does not feel like a gimmick. It does feel like a genuine shift in the architecture of knowledge. In the ongoing quest to reduce the friction of inquiry, perhaps a leap as consequential as a spreadsheet but for a different mode of analysis—words, not numbers. And the stakes here are moderately high: as the digital commons floods with noise, can context-aware assistants truly deliver clarity?
So here is the state of things ten days after going all-in on The Browser Company’s Dia-AI ChatBot-turbocharged web browser: <http://diabrowser.com>, as I said I would.
You may (or may not) recall that my objectives are to:
My major reactions so far: Principally, having a natural-language right-hand pane in a browser window—one that you can interrogate at will, and that returns (mostly) informed answers in real time—strikes me as a paradigmatic instance of what, over the past three-quarters of a century, has made computers so transformative: the progressive tightening of the feedback loop between human intention and machine response. This is not merely a matter of convenience; it is a qualitative shift in how we engage with information. When, in the 1970s, VisiCalc first allowed accountants to see the immediate consequences of a changed cell in a spreadsheet, it revolutionized business practice by collapsing the latency between question and answer. The same logic underpins the success of Google’s search box, which—at its best—turns idle curiosity into near-instant knowledge. Now, with AI-augmented browser panes, we are witnessing the next iteration: a context-aware, conversational interface that not only fetches but interprets, summarizes, and, on occasion, even critiques. Thus the true engine of technological progress has always been this relentless drive to reduce the friction of inquiry, to bring the world’s knowledge ever closer to the point of use. Such tools do not simply save time; they reshape what it means to know, to learn, and to decide. Dia Browser’s WebChatBot windows are an attempt to push this to the max. They also generate AI-slop. There is a tradeoff here. A powerful thing making this tradeoff potentially advantageous—and this applies not just to Dia, but to an entire new generation of browser-embedded AI assistants—is their automatic “awareness”of both your immediate context in the web browser tab that currently has focus and your accumulated digital preferences and browsing history. This is merely a matter of convenience. But it is a very powerful convenience, both in potential and to a substantial degree right now. The browser can now become a semi-active collaborator, not a completely passive conduit. Perhaps the relevant analogy is that Dia is to web browsing much as Cursor has become for programmers—serving as a context-sensitive copilot, debugger, and assistant within the codebase. Programmers have long benefited from IDEs. Now Cursor, because it is aware of that IDE context, goes far toward. understand their logic, anticipate their needs, and suggest improvements. With tools like Dia, the rest of us may enjoy a similar leap in productivity and insight as we navigate the web’s vast and unruly information space. How does it do what it does? Well, its guardrails keep it from revealing its actual system prompt. But it does have views on what a best-practice system prompt would be for a browsing-assistant ChatBot. It is this:
The question is: does its kernel-smoothing attention mechanism pick up its actual system prompt as the most relevant thing it has in its total context to the question of designing a best-practice hypothetical system prompt? It would seem to me likely that this is so. But: It still cannot reliably deliver a University of Chicago author-date format for an article or a book, or a working link to the book at <http://archive.org>. And as for the WebChatBot:
(2) If Dia Browser’s WebChatBot has, so far, slightly overshot my expectations as to its path toward usefulness, the rest of “AI” has somewhat undershot my expectations so far this year. And now we have the release of ChatGPT5. First, Casey Newton comments:
But he then bulls ahead anyway:
Plus we have:
The most obvious first point: This is not what we were told would come with GPT5. We were told GPT5 would be a meaningful step toward AGI. This ain’t it. This is a lot less “close” to anything properly Turing class. Stepping back, OpenAI has repeatedly been building models, deciding that they were not worth the “GPT5” label, and trying again. Now the marketing imperative seems to have overwhelmed the “we have to show that we did not overpromise” desire. Why is this not “GPT4.7”? I do not know. Second, there may be a broader pattern here. On their HardFork podcast, Kevin Roose joins Casey Newton and they both start before expressing what I see as measured disappointment with ChatGPT-5,
before going on to roast Amazon’s Alexa+. After trying the new Alexa+, they apologize to John Gianandrea, Robbie Walker, and the entire Apple executive team. They apologize for all the mean things that the HardFork podcasters said about them when they failed to release the upgraded Apple Siri on time. Apple was unwilling to ship a still-broken voice assistant. Roose and Newton harshly condemned them for it. But now, having seen Amazon’s release of its own half-baked Alexa+, they recant:
Apple’s restraint, in this instance, is now seen so much an abdication as a recognition that, in the arc of technological history, credibility is built not by the volume of releases but by the substance of what is delivered—and that while Amazon can ship something broken and let the chips fall, Apple has to deliver things that “just work”. Third, my feed was overwhelmed on GPT5 release day by people—Ernie Svenson, Gary Marcus, Dave Winer, Ethan Mollick, Eric Newcomer, Simon Willison, Azeem Azhar, Tyler Cowen, and a host of others—who had instant opinions on GPT5 within hours of its release. They had them even though they knew as well as I know and Casey Newton knows that “the worst day to review a model is the day it comes out: there's simply too much that you don't know”. Falling into this “immediacy trap” of hot-take internet intellectualism is not a virtue, and erodes your salience. Takes are worth offering only when there is a chance that they are truly worthwhile Fourth, I now dismiss those who keep insisting to me that one cannot truly grasp the utility of “AI” models without investing considerable time in mapping the “fractal” contours of their capabilities. But, here, once again, as last year and the year before, such metis-type knowledge of how the tool actually works is evanescent. It is not so much that the models get better, it is that they get different, and so rather than building expertise, yesterday’s insights into prompt engineering or model idiosyncrasies become mere archaeological curiosities. What you have, in effect, is an internet s***poster in your pocket—an omnipresent companion whose default outputs echo the cacophony of the digital commons as to illuminate it. The challenge is to prod and cajole this stochastic parrot until it hill-climbs, through iterative prompting, to a region of its vast training data space where information becomes reliable, intelligent, and concise. You have to offer query and correction nudges to get the model away from the noisy valleys of meme culture and toward the higher ground of substantive and knowledeable discourse. The promise of having “expert-level intelligence” on tap is thus always mediated by the necessity of skillful engagement—but skillful engagement that you cannot practice very well, because what worked last year may well not work this year. Thus I suspect that in the end it will prove far more effective to deploy a smaller, more nimble model—an on-device model—one that excels at the natural-language interface part, the “transformer” as a conversational front end, while delegating the actual retrieval of reliable information to some structured database or a curated, trusted corpus of text and data. This, after all, is the architecture that has long underpinned successful information systems: the division of labor between a user-friendly interface and a robust, verifiable backend. The virtue of such a hybrid approach is twofold: it not only enhances reliability by tethering the model’s output to trusted sources, rather than the collection of all internet s***posting. It also cultivates a salutary humility—an ability to admit failure, to say “I don’t know,” rather than hallucinating plausible nonsense. That would mean that the truly productive labor should shift toward the frontier of high-dimensional, big-data-driven, and flexibly parameterized classification and prediction—leveraging the full power of advanced statistical learning on a well-curated data corpus. This is, I think, where the transformative promise of machine intelligence lies: not in the superficial mimicry of conversation, but in the capacity to discern subtle patterns, forecast outcomes, and generate actionable insights from oceans of information. The expanding ability to process and interpret ever-larger and more complex datasets is where we have been going from Hollerith’s punch cards to the neural nets of today. Real progress has consistently come from the ability to extract actionable knowledge from complexity, whether through the invention of the spreadsheet, the rise of econometrics, the deployment of statistical learning algorithms, and so forth. But the AI-industry is unlikely to go there, because that cannot be easily sold as steps to AGI or ASI, can it? If reading this gets you Value Above Replacement, then become a free subscriber to this newsletter. And forward it! And if your VAR from this newsletter is in the three digits or more each year, please become a paid subscriber! I am trying to make you readers—and myself—smarter. Please tell me if I succeed, or how I fail…#mamlms |
| ||||||||||
No comments:
Post a Comment