wnmurphy

It's funny to me that these models were created by effectively "distilling" all available content including the proprietary works of many other people, but now it's a problem that someone is doing the same to them.

You're using available information (copyrighted works, or the output of another model) to train a model to encode the information in a new form. Why is the former not theft, but the latter is theft?

I think the author's choice of words is framing the discussion. They did build their own website, but they loved the look of the one they saw, so I'd think a better choice would be "inspired by" rather than "stolen from."

This was my first thought as well. Hunter S. Thompson used to copy Hemingway by hand to internalize his cadence.

The telemetry will tell the story, not the clickbait news headline based on preliminary information. I seriously doubt FSD would drive into a house.

Until Tesla (or someone else) actually takes responsibility while the car drives itself

Were you aware they're already doing exactly this in Austin and Dallas?

Another factor I don't see mentioned: every Trump policy is inflationary. Deporting the labor base, deterring legal immigration, applying tariffs that were basically derived by taking the trade deficit with a country and dividing by 2, and now locking up 20% of the global oil market by shutting down the Strait of Hormuz.

The Fed's response will have to be to raise rates, which is going to crush the multiple of any stock whose valuation is based on the expectation of massive growth.

FSD is up to v14 and there is no sense in which it's "Full", or is likely to be "Full" any time soon.

I think the SpaceX IPO is way overvalued, but have you actually tried FSD v14 yourself?

My car has 8-year old hardware running v12, and it handles like 90+% of my driving. When I test drove 14 it blew my mind how good it had gotten in only 8 months of development. In my opinion, there's question that it's "when," not "if."

Which means more Grok degradation, more severe throttling, etc.

I can't understand why xAI charges 50% more per month for Grok over competitors when it doesn't even gracefully downgrade to a cheaper model when paid subscribers hit the limit.

This is the exact analogy I use to explain how models work to laypersons as well.

I'm considering Anthropic. I think they will be one of the survivors if/when the AI bubble bursts.

I was dubious about SpaceX (orbital data centers need to solve for extreme radiation and error-correction during training), but then I remembered that xAI is actively working on virtualizing white collar workers ("Macrohard").

In my opinion, this is the only TAM that justifies $1T in data center investment, because the consumer market for ChatGPT-style AI is saturated. There's a lot of enterprise TAM available for AI, but I think what these companies training frontier models are really after is selling a product that allows companies to eliminate the cost of white collar salaries.

I really want Spotify to follow. I feel cheated and deceived when I'm enjoying some music, then I realize that there's no bio for the artist and they released 7 albums in 2025. Users should be empowered to filter out AI content if they choose.

A while back, I discovered by accident that forcing ChatGPT to respond only in zalgotext and then using voice mode generated some really bizarre audio.

Have uploaded the audio files for anyone who's interested. They've since patched the ability to ask for output in zalgotext, but it was pretty strange while it lasted.

My car drives itself. That's a $18T global market.

Also $1T in data center investment makes sense when you realize that companies are racing to create virtual white collar workers. Google spends $9B a year on software engineers.

Exactly. It's a tangent, but clearly a pain point for enough users.

100%. I have to hold the floor by filling the space with "ummmmmmmm.... uhhhh...." which inevitably distracts me from my point altogether. Poor user experience.

Strongly agree, some of us like to choose our words more carefully when interacting with an LLM.

I've tried to convey this to OpenAI through various available channels (dev forums, app feedback, etc.).

Grok solves this by having an optional push-to-talk mode, but this is not hands-free and thus more cumbersome than just having a user-configurable variable like seconds_delay_before_sending_voice_input.

I agree with your categories. The majority of the usage for me is (1) and (3).

(1) LLMs are basically Stack Overflow on steroids. No need to go look up examples or read the documentation in most cases, spit out a mostly working starting point.

(3) Learning. Ramping up on an unfamiliar project by asking Antigravity questions is really useful.

I do think it makes devs faster, in that it takes less time to do these two things. But you're running into the 80% of the job that does not involve writing code, especially at a larger company.

In theory, this should allow a company to do more with fewer devs, but in reality it just means that these two activities become easier, and the 80% is still the bottleneck.

Tangential, but you used to be able to use custom instructions for ChatGPT to respond only in zalgotext and it would have insane results in voice mode. Each voice was a different kind of insane. I was able to get some voices to curse or spit out Mint Mobile commercials.

Then they changed the architecture so voice mode bypasses custom instructions entirely, which was really unfortunate. I had to unsubscribe, because walking and talking was the killer feature and now it's like you're speaking to a Gen Z influencer or something.

Not saying that the stock isn't a meme stock, but my car literally drives itself everywhere. Tesla has many business models.

I heard he used to be _The_ Whitney Brown.

Our understanding of the world is overfit to the macro level, where we project concepts onto experience to create the illusion of discrete objects, which is evolutionally beneficial.

However, at the quantum level, identity is not bound to space or time. When you split a photon into an entangled pair, those "two" photons are still identical. It's a bit like slicing a flatworm into two parts, which then yields (we think) two separate new flatworms... but they're actually still the same flatworm.

Experiments like this are surprising precisely because they break our assumption that identity is bound to a discrete object, which is located at a single space, at a single time.

Essentially all intelligent life is a pachinko machine that takes a bunch of sensory inputs, bounces electricity around a number of neurons, and eventually lands them as actions, which further affect sensory inputs.

This metaphor of the pachinko machine (or Plinko game) is exactly how I explain LLMs/ML to laypersons. The process of training is the act of discovering through trial and error the right settings for each peg on the board, in order to consistently get the ball to land in the right spot-ish.

I recognized the word "glymphatic" from recent articles about the discovery of the brain's self-cleaning system, and then understood from the headline that these authors identified that the mechanism by which this occurs is driven by norepinephrine.

I know how to plug this black box into this other black box and return the result as JSON!

To be fair, most of software engineering is this.

At some point, we will have no idea that the majority of the commenters we're interacting with are actually just generative AI.

Related: I've found that the internet becomes significantly better when I use a Chrome extension to hide all comment sections. Comments are by far the most significant source of toxicity.

Introduce intermediate variables with meaningful names

Abstracting chunks of compound conditionals into easy-to-read variables is one of my favorite techniques. Underrated.

isValid = val > someConstant

isAllowed = condition2 || condition3

isSecure = condition4 && !condition5

if isValid && isAllowed && isSecure { //...

The very first sentence is highly biased and draws an invalid conclusion. They attribute this to vehicle demand, without basis.

It's likely due to retooling for a planned lower cost trim.

Elektrek mostly writes misleading articles like this.

I think the argument is in whether "thought" only applies to conscious articulation or whether non-linguistic, non-symbolic processes also qualify.

We only consciously "know" something when we represent it with symbols. There are also unconscious processes that some would consider "thought", like driving a car safely without thinking about what you're doing, but I wouldn't consider those thoughts.

I find an interesting parallel to Chain of Thought techniques with LLMs. I personally don't (consciously) know what I think until I articulate it.

To me this is similar to giving an LLM space to print out intermediary thoughts, like a JSON array of strings. Language is our programming language, in a sense. Without representing something in a word/concept, it doesn't exist.

"Ich vermute, dass wir nur sehen, was wir kennen." - Nietzsche, where "know" refers to labeling something by projecting a concept/word onto it.

100%. Test names should include the word "should" and "when". Then you get a description of the expected behavior.

I'm always amazed at how many nurses take your BP with your legs hanging off of the side of a table, back unsupported.