I think this is what people mean when they say LLMs are a higher level abstraction. We still need to consider edge cases and have tests. We still to sweat the architecture and understand how the pieces fit together and have a mental map of the codebase. But within each bottom node of that architecture we don't sweat the details. Anything obvious gets caught right away. Most subtle/interaction-based issues occur at the architecture level. Anything that bypasses those filters is a weird bug that is no worse or different from a normal bug fixes - an edge case that was hit in a real world scenario that gets flagged by a user or a logged as an error.
There are certain codebases and pieces of code we definitely want every line to be reasoned and understood. But like his API endpoint example, no reason to fuss with the boilerplate.
This has definitely been my shift over the past few months, and the advantage is I can spend much more time and energy on getting the code architecture just right, which automatically prevents most of the subtle bugs that has people wringing their hands. The new bar is architecting code to be defined as well as an API endpoint->service structure so you can rely on LLMs to paint by numbers for new features/logic.
I remember around 2000 I read about how Ted Turner started his empire: he bought podunk local TV stations that had loose contracts with media owners that allowed them to broadcast shows as often as they wanted, with no restrictions. In the those days, local TV stations were broadcast just like radio and so the assumption was the contract only concerned the audience the TV station's antenna could reach. But the contract didn't specify this. Recognizing the loophole, he bought multiple stations and combined that content into its own cable channel(s) that played old movies and TV shows: https://en.wikipedia.org/wiki/Ted_Turner This was the basis that allowed him to branch into CNN and more.
When I learned about this, the story was very applicable to me at the time, as my startup had acquired licenses for content that was historically sold directly to libraries by a salesman who would negotiate with each library individually. He used a standard contract. When we contacted the company to license content for display on the internet, they gave us a ridiculous contract with a small one time fee and access to display the content forever. Only after reasoning through their business model and history did we understand how this occurred, which was exactly the same type of gap that Ted Turner had exploited.
Also if you're in the Boston/Cambridge area, WMBR is a fun and weird listen and clearly sounds like college radio when compared to something like WERS.
I was wondering if they are going to put Ted's crayons in the box with him. At the time of this first being done it was so comically bad, and the jokes were ruthless. As much as I'm not a fan, the modern AI stuff is so much better without saying it's good. That's just how bad Turner's colorization was. The best colorization was Weta's footage from WWI where they used the actual uniforms in the images as reference rather than just someone adding color based on the feels.
When my nephews were kids I used those old colorized movies from Turner Classics as partial proof that the old joke about the world being black, white, and shades of grey when I was kid was true. They grew up in the late 80's and early 90's watching TV shows including some great old stuff that Turner later colorized. I had told him how scientists had discovered how to improve the appearance of everything by adding other colors and as a result, scientists and artists and representatives from around the world met and collaborated on methods of colorizing everything that existed. Everyone agreed that blues would be great for the sky to lighten things up after storms; animals needed fur that blended into their environment so browns and tans like the dirt outside; rocks could be any color but earth tones (like their Mom was using in painting their house) got their names after everyone had picked colors for rocks, tree bark, leaves, etc. Plants would be green for the most part but leaves that had lightened or darkened in the fall could change colors too so every continent and country was able to decide how to color flowers and plants as they wished since coloring all flowers one color would just be boring. Snow and ice were white and water was up for grabs especially if it was in a river.
The notes they could read in the movie credits about it being a colorized version simply told them that all of the colors in that movie had been added later.
I was so convincing that one of them interrupted his teacher in class to let her know she was wrong about the rainbows and where color came from. I had made it clear that everything that we saw as colored had the colors that were assigned by international agreement after people had become tired enough of the BWG palette to sit down and make it all change.
In the end, the teacher told him he was wrong and he argued about it so I got a call one day that he had been in trouble at school and that the teacher was not thrilled to hear his explanation so I needed to clear things up for him since he was not inclined to believe her at all. I'm not sure that I ever got that completely cleared up because, to me, it was just too funny that I was the most trusted source.
Just wait until we simulate old films and media and turn them into living, breathing VR games.
We'll eventually do that for all of history. At least the history we have samples of or can plausibly recreate.
I'd imagine playing one of those might be like living your life right now. Punctuated by lots of mundane, lifelike moments.
Like reading an "internet forum" full of other period appropriate "humans".
Sounds a bit dystopian where people are living in what amount to a synthetic reality. At that point reality’s importance decreases and at some point the virtual reality becomes the more normal existence and people eschew actual reality. And at some point you think why not virtualize all experience and then life itself…
I kind of wonder if there were color photos of the actors and scenes from the time of some of the black-and-white movies. You could use them as conversion-training-data with AI to auto-colorize the movies.
Rightfully so if you ask me. Out the gate think about the implications of determining, say, skin color. I’m not saying “under no circumstances should it be done” but I also think people don’t appreciate the importance of the decisions made and the politics/implicit biases under the hood. I’m not even getting in to artistic intent and impact on lighting here either.
Colorizing b&w images is still debated to this day.
Because of the film technology at the time, a lot of the skin tones on set wouldn't match what you'd expect anyway due to makeup designed for the b&w film. Lots of sickly greens, yellows, and blues in place of red tones for instance.
At that point if you've already decided you want to colorize the film, there's a real question of how do you approach it, because being true to what was on set definitely isn't the right choice. So now you're playing with skin tones regardless.
Huh. That actually brings up a kind of modern parallel I hadn't thought of. A lot of action movies are done primarily, or in part, on greenscreen. The intent of using a greenscreen has nothing to do with what was captured, and more so to do with what is trying to be depicted; what ought be seen, not what is being seen by the actors and actresses.
It would be interesting to know if, in say, 100-200 years, there is some alternative technology that could de-render todays CGI perfectly, and then replace it with some alternative, perhaps insert some form of practical effect in a convincing way? Would being able to do so be better to do just because it can be done?
Like, suppose that one of the more recent big budget movies, Transformers or whatever, could entirely have all of the CGI stripped out of them instantly, and then be replaced with some form of "less fake" effects in a different way. Would it be good to do so, if that were possible? For me personally, I'm very much in favor of rubber suits and fake blood over sticks with ping pong ball overlayed with graphics. [1] In spite of my preference though, I don't know if however many hundreds of people who had worked the digital modeling for all of those scenes would appreciate essentially deleting all of the thousands of hours they had put into the movie.
Bringing that back to B&W films, I think that if someone was really excellent at doing the set design for B&W films, it makes me wonder how they might react if someone insisted on "fixing" the film by colorizing it, and showing their set pieces in a way that they never intended for those pieces to be seen by the audience. Like, if they weren't outright upset with even the idea of doing it at all, perhaps they might insist on some sort of creative control on how each of those set pieces were colorized and portrayed in the final product. Obviously, that would then extend out to all of the other things too, like wardrobe, makeup, etc. I could see the complexity ballooning out to be as complicated and involved as making the movie was to begin with! For example, maybe the guy that scouted the original location for the film wouldn't have chose the spots he had chosen if he knew that people would be able to see it on giant TVs that they could pause every single frame of, and perform all kinds of upscaling and digital zooms in and out on.
[1] I am firmly in favor of practical effects over digital for everything, except small technical errors like a boom mic or a coffee cup in a shot, because I think that the constraints a movie set faces will demand either: incredible innovative solutions by the crew, or, those constraints force directors to scale their vision back to something more contained and manageable. It helps to show where the scope creep for a movie is, and where it's simply unnecessary. For example, Jaws has a great backstory regarding the constant issues of the mechanical shark, it really forced Spielberg to rethink how and when the shark would be shown, and when it would be better to let the viewers mind fill in the blanks.
I think these are really interesting questions and I like a lot of what you’re saying. I don’t really agree with your near prohibition on CG, but I definitely get where it comes from and think that some productions definitely abuse it
Eh regarding skin color people don’t care about realism these days. You have historical remakes with totally anachronistic ethnicities in them and “no one” cares.
I mean sure, some people do, the same as some people used to complain about overrepresentation of caucasians in some old movies set in what was then called “the orient”. I think the only ones who put up a fight are the Japanese who don’t like their productions ethnically misrepresented as much.
B&W highlights the stories better. With color you get more ambient context and sometimes that’s interesting.
I think you have a misperception of the past. The actors that played the great chinese detective Charlie Chan were Warner Oland, Sidney Toler, and Roland Winters.
> Eh regarding skin color people don’t care about realism these days. You have historical remakes with totally anachronistic ethnicities in them and “no one” cares.
This isn't exactly the same thing. Colorizing historical footage decides what the color is. A remake is an interpretation with nowhere near the same claim of accuracy and the audience 100% knows this. The social politics of this are incredibly important.
I had a chuckle at your comment and felt it was true. But wonder if the commenter is younger. Ted Turner was much more of a household name and public figure in the 20th century. He became less involved in the cable empire by the mid 90s. Younger millennials and onwards probably heard people talk about him a lot less.
Ps. Another memorable media portrayal of Turner, he was clearly the basis for the boss character in the 1994 cartoon The Critic.
I don't think anyone else with the name Obama has been president of anything that confers a library (let alone a presidential library), your answer seems a bit needlessly derisive. I suspect you're just insecure about your personal level of useful knowledge and are trying to lord over someone with your trivia fact.
Don't undersell AI - it also synthesizes and recombines those summaries in a purposeful way. Otherwise it couldn't product code that works in an existing codebase.
So it is able to process and act upon summaries and concepts. In other words, apply synthesis. What it can't do is understand what a useful result looks like without direction. So it could synthesize a billion pointless claims from source material, but we still need a human to know which ones matter (without a specialized framework to comprehend this). If you provide LLMs with an objective and source materials it is certainly capable of following threads of logic or building an argument backed by sources.
I understand the concerns about AI, but it is a powerful tool for discovery and synthesis.
Another thing they’re often poor at is making an incorrect assumption and then going down a rabbit hole trying to unnecessarily solve for it. Without a discerning human in the loop, you can end up with large amounts of unnecessary output.
Like the standard of denying people entry to the country based on their social media posts? Or deporting them for the same? Or the standard of tear-gassing a peaceful "No Kings" crowd of U.S. citizens, full of children?
Those, to me, are mainly authoritarian tendencies of the current administration.
That is a different argument: The Trump administration is not really shifting the defamation vs. free speech tradeoff in the US (you could argue that it does, in the opposite direction, by slandering political opponents with insulting nicknames like "crooked Hillary" or "sleepy Joe").
100%. It's important to realize our understanding of "dark matter" is fuzzy because we only understand it through data anomalies. Dark matter is a classic catch-all concept that we use as a crutch while we try to understand the underlying system better.
Similar to how we used to believe in "aether" to explain how light could travel through empty space. It is important to understand how these crutches help and hinder understanding.
It's kind of weird in this case, though. All the math acts like there's something invisible and heavy everywhere that we find clumps of visible matter. When we look at the motion of galaxies, they behave as if they're much more massive than the count of stars and such in the would have you believe, and in ways that otherwise jibe with our understanding of physics if only that galaxy were heavier.
If you have one galaxy that's acting heavier than you can eyeball, measured by things like light bending around it, then maybe you have some weird phenomenon. When every galaxy calculates out to be about 6x fatter than you'd expect, something else is going on.
I think we are wending toward a solution here for context, because no matter how big a context window is, there needs to be a way to navigate and prioritize that context, a way to handle contadictory info, etc.
So we need a taxonomy, we need memory layers, we need summary/details. If there is one thing I have learned about how these LLMs work, if you give them a few flexible tools they can work the shit out of them to achieve objectives. We just need to right tools and right structure for context.
I did something similar, but instead of having the LLM play the game I had it build an entire bot system to play the game. Bots require much more determinism, but I'd rather burn tokens encoding problem solving approaches and bot decision profiles than using LLMs for every turn of the game. This can be developed rapidly if you create an agent in a loop and say "figure out how to have the bot reach room 3 in under 10 actions" or something like that. It is easy for this to get bloated, but I found it makes a nice feedback loop that allows me to quickly test things like pacing changes and think of the game as a series of user actions that can be sculpted purposefully.
There is a ton of optimization possible when we are able to observe how LLMs and agents process and navigate our code given different prompts. For example, our MCP was pulling down way too much data to resolve a simple "count rows" request. Once you see it, it's easy to resolve but I don't know of a good framework yet for walking through some of these patterns.
I built an eval framework to look just at tool calls given a static prompt, with the idea that LLMs should be able to deduce the best tool calls and arguments needed to get requested data. Not as great as full observability, but helpful for complex tool interactions. Anyone have any good tools for this problem?
In the same way we mentally walk through deterministic logic, SWEs need to learn to anticipate LLM context and tool awareness, which is much trickier to reason through, especially given the various LLM IDEs and how they manage context as a black box.
Apparently this doom marketing strategy is working for landing enterprise deals, but boy these AI companies are stirring up consumer hate and fear.
I think the real purpose of the Mythos security sham is to mask that Anthropic simply can't release their new model because their data centers are already on fire. There are so many other red flags pointing to this: the no-Claude-Code-for-Pro-users "test", the AWS data center rental deal, the fact Microsoft rug pulled hard on Copilot, specifically removing Opus... and that's just the past 2 days?
There are certain codebases and pieces of code we definitely want every line to be reasoned and understood. But like his API endpoint example, no reason to fuss with the boilerplate.
This has definitely been my shift over the past few months, and the advantage is I can spend much more time and energy on getting the code architecture just right, which automatically prevents most of the subtle bugs that has people wringing their hands. The new bar is architecting code to be defined as well as an API endpoint->service structure so you can rely on LLMs to paint by numbers for new features/logic.
reply