Context: I did not study CS in college & wished I had so I read the python data science handbook, K&R's C, and Stroustrup's Programming Practice and Principles. I spend a significant amount of time bashing my head against getting code for personal projects to work (it usually did, eventually!).
I am way way way more capable with AI agents for this experience than if I had not invested the time. I can look at the code they wrote & decide if it's good enough or a bad idea. I have instincts about how to do things. I am much more able to plan things with them. My coworkers with degrees are wildly productive.
If you think you will be interacting with code in any significant way, I would at least read those three books, they will help you understand "how computers really work". If you want to have a career in making software, then definitely learn to write programs without AI.
I think what you've said here is unfair and overly cynical; nowhere have I read Chris or any of the Anthropic people make the claim that LLMs definitively are conscious. What they say is that there is increasing uncertainty, and evidence - for which they show receipts - backing up that uncertainty.
Personally, I have been unconvinced by any of these definitive yes or no answers to "are LLMs conscious?" - in both directions. The "they are not / can not be conscious" side relies too heavily on mechanistic reductionist arguments that can apply equally to neurochemical processes in the human brain - and yet, it seems that humans, with brains made of these neurochemical processes, are conscious. At the same time, the "they are definitely conscious" answers seem to generally rely too heavily on self-deception and lack of reality testing.
To be able to say something definitive here, it seems to me that one would need to say definitively what this experience is. I have not heard any of the loud voices in the arena - not Chiang, not Giulio Tonini, not Karl Friston, etc - do so. Therefore I find Anthropic's uncertainty, and careful, caring investigative process, well grounded given the evidence.
My aim isn't cynicism so much as an interest in clear an evidence based information about these systems which are clearly important technological developments for the future. It may turn out that there is something analogous to cognition going on and that would be quite significant and require us to think differently about the technology, but we need to be very sure of something like that.
> Therefore I find Anthropic's uncertainty, and careful, caring investigative process, well grounded given the evidence.
I'm inclined to agree mostly, but from the outside its impossible to separate marketing, belief, scientific inquiry, and just plain enthusiasm. Its all to opaque.
And it is worth noting, I think, that Olah's message to the pope is centrally making that point: that Olah and his peers are within incentive structures that distort their actions and understanding, and therefore that dialog with those outside of such incentive structures is necessary (not sufficient) for their outputs to be trustworthy.
I really wantt to get into splatting and I have the tools: good camera, v comfy in blender, comfy with graphics programming ideas, 4080. But I haven't found a good 'all in one intro' to it yet. Possibly because I'm foss-biased and have dismissed proprietary options. But does anyone know of a good 'vertical tutorial' on this stuff?
I recently got into splatting. I looked for some good all-in-one tutorials, but didn't find any, and mostly muddled through through trial and error and LLM assistance. I present this workflow as a straight-line pipeline, though in practice it took a lot of iteration and backtracking and rework to get the final result. Here's what worked for me:
I captured a video on a smartphone camera, using the OpenCamera app. Specifically, this video was captured with exposure locked, framerate locked, focus locked, fairly high framerate and resolution. I walked slowly and carefully around an outdoor scene, trying to get fairly good coverage from multiple angles. I took roughly 20 minutes of video, weighing 19GB.
This video was sampled into individual image frames at about 5fps using ffmpeg. There's room for experimentation and improvement here, an adaptive, coverage-aware sampling strategy would be better. But fixed 5fps was Good Enough (tm). This resulted in roughly 8,000 images at 4k. This was a pretty hefty dataset for my limited 1080, but I made it work.
I then generated masks for these images, to ignore transient objects during the splat training. (i.e. to cut out people who transiently walked through the scene). For this I used Cutie (https://github.com/hkchengrex/Cutie). For outdoor scenes, it can also make sense to mask out low-parallax areas like faraway mountains or especially the sky, as these are difficult to train correctly. If masks are generated for some images, you'll need at least placeholder masks for the all of them. In the end I've got about 8,000 PNGs that are monochrome black/white masks.
Then the images are handed to COLMAP (https://github.com/colmap/colmap), using the 'global mapper' option. This registers the camera positions in 3D space, and generates a crude point cloud that's good for sanity-checking. This step required a fair bit of iteration to get right. The full reconstructed output from COLMAP is not necessary, only the pose-estimate .bin files. The output directory here was about 500MB for this step for me.
With COLMAP registration done, the next step is the actual training. I found two useful pieces of software for this, with different tradeoffs.
Brush (https://github.com/ArthurBrussee/brush). Was very straightforward to install and use, requiring very little in external dependencies and setup. It was also pretty speedy on training, and gave good results. Minor modifications to the training process were possible by editing source, though I didn't get too wild here. Brush takes the *.bin files from COLMAP, plus the original images directory, and the masks directory if it exists. Run on its own, this could produce gaussian splat .ply files, 500-800MB in size, containing 1-10M splats. More than that and my poor little 8GB of VRAM OOM'd.
nerfstudio (https://github.com/nerfstudio-project/nerfstudio) Was also useful, as many research papers get implemented in its framework. In particular, for this outdoor scene, I used wild-gaussians (https://github.com/jkulhanek/wild-gaussians/) to generate just a sky sphere (to help seed low-parallax areas in my particular dataset), stopped training, and used this as an init.ply to pass to brush.
I then set up a very simple viewer website, using SuperSplat (https://github.com/playcanvas/supersplat). I used supersplat's editor to align the splat's coordinate system with the rotation and scaling that I wanted, and then exported an optimized .sog file, roughly 1/10th the size. .sog is nominally open-standards, though I'm not aware of any other projects using the format. This gave fairly good framerates and adequate controls across a variety of platforms.
As a little bit extra, supersplat's splat-transform CLI tool was used to generate a crude collision mesh for the scene, enabling a walking mode that respected object boundaries.
If there's interest I can post my results, I got a bit sidetracked with other projects and other splats, and this particular one I got fiddling with some more cleanup. I can get it up with a few more hours work. But hopefully that's a good start, all of these are fully FOSS, and resulted in a good-looking splat.
Haha, fair enough. Perhaps a more insightful take from me is that the app makes you understand the underlying process a little more than many other options. For example, color space, color correcting. It gets complicated quickly and since the internet is the source of most of the guidance for noobs about it, there are opinions.
This is a fair criticism. I was lucky enough to grow up on blender and spend lots of free time with it, so all of this stuff is second-nature to me. I am not the standard case, but when software insists on holding my hand and abstracting stuff away it feels patronizing and frustrates me so much I will change my life to avoid using the software.
I tend to let Feynman, Fermi, and Bayes guide my inquiries generally in that order. Part of the process is generating good questions; another part of the process is generating good actionable questions (tailored to the moment). Questions which are obvious, which are "horseshoes and hand grenades" type questions, based on what we know now. Follow things a little bit, test it, see where it's soft.
In this case I called public health and building inspection agencies and asked them what sort of ongoing inspections there were for clinics and other medical facilities. That turned up databases online, and keywords which turned up other databases.
What was the germ (pun intended) of this inquiry? Several years earlier, sitting in the waiting room of a different clinic, and the linen supply company pushes a cart through (gets buzzed through to the back) to collect the dirty linen, wearing gloves. Like they did this every day. Many years ago, the memo taped to the doors to a lab wing at a biotech: "gloves must be removed when greeting visitors". Various reports over the years concerning improperly sterilized dental instruments.
Tangential, but where do you draw the line on commercial advertisement? If a podcast is sponsored by a business & supports behavior the business benefits from, is the podcast advertising? What if it also contains useful educational content related to that activity?
There probably will never be a clear line and none of this is realistic, but I would start with banning all flashy light polluting physical advertisements. Any advertisement people cannot evade.
A podcast I can choose to listen or not.
A news site I can also evade. But any (internet) service people must use, should be ad free. Ideally all of society, but any regulation here will have a hard time in the real world.
Pretty decent, it's given similar book recommendations as Claude when I feed it my list of read books and thoughts on them. You'll have to tell them to never use emojis. I was using 3.5 a while ago to generate some flavor text while I was playing a bit of an old-school dungeon-crawler game (it's like Wizardry), a genre I don't particularly enjoy much, but it's funner with the flavor text. Worth setting up something like open webui or other front-ends since a pure CLI experience via ollama is pretty bad.
If you're in a country where war is occurring, it doesn't matter if it's a world war or not. There are conflicts in pretty much every continent. North America is waging war in the Middle East. Europe has a multiyear conflict threatening to spill across more borders. Several countries in Africa are in conflict even if they are civil wars. North America, while not waging outright war, is in conflict with a South American country. The Asian continent is nearly routinely going through border skirmishes. Antarctica doesn't count. The Australian continent seems the only one without active conflicts. So 5/6 continents capable of being part of world war is in warlike conditions.
Can anyone give me a semi-technical reason on why the hydrogen division are delusional? I'm actually convinced of it "osmotically", but I just don't know enough about it. I've got chem 101 behind me but otherwise I'm a finance & tech guy. It would be nice to actually understand why it can't be done though.
If we manage to get enough solar such that energy essentially becomes infinite then the inefficiency would no longer matter. Otherwise, it would only make sense in vehicles that require high energy density like airplanes.
There is already free energy. In 2024, California curtailed 3400 GWh of solar. Hydrogen is one of the easier ways to load shift that to winter or processes which need something denser than batteries. I actually prefer synthetic methane (worse efficiency) because it is more immediately usable.
Looking at the charts[0], load shifting all the way to winter seems unnecessary. You only need to load shift until ~6pm, in which case there are plenty of better options out there like grid-scale batteries, flywheels, pumped storage hydro, etc.
It depends on your energy mix. The more solar you go, more load shifting is required. Winter solar production is relatively crummy and you need to offset that loss somehow for the entire season.
That chart is showing some curtailment in winter because the grid knows to expect less production. It is already tuned to spin up more gas because solar will underperform relative to Summer.
Hydrogen leaks too much. You cannot transfer it without spilling a good chunk, around 8 percentage of it into the atmosphere. Now match that loss by your long term refuel needs and you'll simply run out of it. It can't scale up to anywhere near what would be considered a replacement technology for EVs.
I am way way way more capable with AI agents for this experience than if I had not invested the time. I can look at the code they wrote & decide if it's good enough or a bad idea. I have instincts about how to do things. I am much more able to plan things with them. My coworkers with degrees are wildly productive.
If you think you will be interacting with code in any significant way, I would at least read those three books, they will help you understand "how computers really work". If you want to have a career in making software, then definitely learn to write programs without AI.
reply