textninja

That’s a moot question because it will continue to evolve, and while its successor will inevitably dwarf its capabilities, it will be remembered as an early influence and historically significant innovation.

If it was frozen in current state, however, it would still be a springboard for innovation because businesses will find new ways to integrate it with their products and compose its functionality with other code and itself. It’s impossible to imagine the kinds of influence these products will have because progress will be following the exponential part of an S curve for some time to come.

Finally, even if in this artificially constrained world of our imagination we don’t allow for third party products that use it as a platform, it will completely upend the way business is done, on a 10 year timescale, just as search engines have before it.

At first, maybe, but then human to human conversation will drift to gated communities with strong verification, and on the wider net it will be like the blogging revolution all over again as people exchange pre-seeded forums for ideas more so than the ideas themselves, and compose them with other constellations of thought using the ecosystem of AI tools built around same.

From where I’m sitting, ChatGPT in its current nascent state is absolutely changing the world. It’s not only showing what’s possible but also making a powerful and highly disruptive tech available to the masses to hack and extend.

Of course, we won’t stop here, but as one of the seeds from which AGI will spring, I truly believe it will be seen as a historical innovation in the same league as early search engines, crypto (yes, unironically), or even the internet itself.

Fair point, but - uno reverse card - maybe it will make tedious exposition so effortless and superfluous that nobody wants to do it anymore.

It might do to overlong text what border-radius did to rounded corners, and discourse will trend back to wit and concision over ceremony and formality. Businesses will value drafts - unpolished bullet points, like you describe - over finished pieces, because they’ll feed everything into standardized models fine-tuned by Brand Prompt Engineers to generate copy in the organization’s voice.

I am an independent contractor, but my comment was in no way intended as a slight to anybody I work with or have in the past. I have had the pleasure of working with some exceptional people.

We may however have different standards for what constitutes doing something “especially well”. I’m talking top percentile performance. That is by definition rare, but that seems to be the standard we’re holding the LLMs to.

It’s a bad take for several reasons.

- It’s reductive. The model may be architecturally relatively simple, at least insofar as humans are involved, but the data that was used train it is anything but. Code is data, data is code. The model is more than its architecture. It’s all the weights and biases within.

- It’s qualitatively wrong. The temperature settings change it from a predictive system (auto complete) to a generative one (digital assistant).

- It’s quantitatively ignorant. How do you eat an elephant? One bite at a time. For that matter, how do you speak if not one word at a time? It happens that the model generates one token at a time as well, but the million dollar question remains: what chain of reasoning did it have to follow to come up with that token? Just because it‘s only outputting a single token doesn’t mean it’s not “thinking” about the structure of the sentence, paragraph, or the entire document - several tokens ahead, to say the least. Indeed, having been trained on a such a large internet corpus it has the capacity to model entire personalities and all the quirks and neuroses that go on behind the scenes.

So, yes, it is a bad take. That being said, it’s still a useful one because it encourages people to think about the models as tools, which is exactly correct given how they’re structured. So I still describe them as such.

> It doesn't really do anything especially well and it's not clear when you should trust it to be correct.

You just described most humans. Imagine someone gave you access to a free digital workforce that you can only interact with through chat; it’s text in, text out, they only do what you say, and although you’re conversing with middle-tier experts they don’t have internet access or even a paper or pen handy so you have to take any facts or hard numbers they reference with a grain of salt. What is the use case for that?

What I’m getting at here is that ChatGPT doesn’t have a single clear use case. OpenAI is positioning themselves to be vendors of foundational AGI and it will be up to the market to create specialized fine-tuned models with higher levels of agency and built-in accountability.

This criticism, like many others, attacks the mechanics of how LLMs think, apparently dismissing the models for not doing so using the same process, faculties, and background life experience as a human. It does not contain any compelling arguments to refute the notion that LLMs think. We may not have consumer level AGI just yet, but suggesting a LLM is just a dumb anything (stochastic parrot or otherwise) is a rather extraordinary claim to make about something that basically patterns the whole internet.

We’ve been here before. Our sense of place in the universe was first upset by the heliocentric model (we’re not as special as we thought), then the theory of general relativity (not as correct as we thought), then quantum mechanics (not even living in a deterministic universe, really). With all these fantastic discoveries behind us, now seems like the right time to learn that we’re not as smart as we thought either.

They did say “new-fangled” though.

It’s useful to make a distinction between what is one’s job and where one spends his time - that is, the theory and practice of job titles. Data cleaning is 0% of a Data Scientist’s job despite occupying 90% of their time. My hope is that AI can help bridge that gap.

I’ve found the criticisms regarding accuracy to be curious, not only for the reason you mention, but also because the apples-to-apples comparison isn’t ChatGPT vs Google so much as it’s ChatGPT vs an expert working on a test in a faraday cage with no substantial penalty for wrong answers.

With that context in mind it’s easy to extract value from it now, and either by allowing ChatGPT to adopt a process, or by building systems around it that enable greater agency (as they seem to be doing with Bing) these issues will eventually be overcome.

Web Components aren’t an alternative to React so much as a complementary technology. Yes, you can build UIs without a reactive framework but it will involve compromises in expressivity, code organization, and whatever the Greenspun equivalent is for React.

I can happily live without TypeScript (ka-chow) but at the same time I don’t see why not.

I think he just means you have to build before deploying and while developing. Even if that build takes fractions of a second and is trivially scaffolded and automated in a modern context, it’s still technically an extra step. There was a time I might have agreed with that take but I’ve since embraced the build since unless we’re talking about the kind of JavaScript you’d embed in a single blog article and then promptly forget about, chances are you’ll want to introduce build tools eventually.

In black and white terms, sure, it’s still a build process, but I second the recommendation to try Vite - the overhead is minimal and the expressive power build chains enable more than justify the overhead. I say this as someone who still writes plenty of vanilla JavaScript and was long reticent to adopt build chains into my workflow.

In this case the A in YAGNI stands for “Are”.

Elixer does seem tailor made for this sort of thing, but I’ve never used it so I can only recommend you try [feathersjs](https://feathersjs.com/). Setting up a chat backend with it was fairly straightforward, but your mileage may vary as it does force some abstractions that you may not want.

I am dying to know the answer to this. I can’t get over the fact people are paying $8 for downloads of randomized triangles, but then again NFTs are a thing so I shouldn’t be too surprised.

Points 1 and 2 are not how to write, but rather how to use lies and rhetorical trickery to showcase your intellectual plumage and profit from deception.

Sometimes, even if it is plain, boring, or obvious, the truth must be told.

I use multiple languages, including Python, and this just reads to me like a hit piece. I won’t take the time to respond point by point, but the author complains about 69 top level functions that “you just have to memorize”? Really? Try learning a foreign language and then come back and tell me again why memorizing 69 function calls is so onerous. Oh, and by the way, you’re not even supposed to memorize them all. Who told you to do that?

Other than that, the piece seems to reveal more about the author’s misunderstanding of language concepts like lazy evaluation than it does to expose failings of the language itself.

Imagine learning a new language and complaining about the “weird” ordering of words or predicates. Weird to who?

Tales of localhost’s demise have been greatly exaggerated.

Putting aside how much this rings of “owning nothing and being happy”, I’m going to go out on a limb here and say no, cloud-based development will not replace localhost, although with the pace of innovation in that space and tangible benefits that follow, developers will naturally begin to lean on cloud-based workflows more and more. That is, up to a point. It being up to a point is key.

The cloud, as an OS, is not peer to an actual OS. Latency, introspection, isolation, and ownership matter, and as long as there are developers - no, humans - there will be a drive to own, with all the layers of virtualization, mocks, faraday cages, and locally configured quirks and quarks that entails.

While I do think it’s here to stay, I don’t consider TypeScript to be “the future of JavaScript”. Given the choice, I’d much rather use JavaScript, and that applies to large codebases as well. TypeScript just gets too verbose and adds way too much ceremony for my taste, the implied safety is skin deep, and I far prefer metaprogramming and reflection to compile time IDE hints. Type systems are not for everyone. See also, Haskell vs Lisp.

Also, I disagree completely with your claim that React is outdated. Plenty of existing projects use it, plenty of new projects are starting with it, it still has a huge community, it’s actively maintained, and this is a subjective point, but the developer experience outclasses pure web components by a lot.

This is very inspiring, thank you! Having never done it, I know first hand how difficult it can be to get your ideas out the door. It looks like you found the perfect tempo and scope, tiny projects rock.