cjf101

I hate to hand anything to Generative AI tools, but

While Great design breaks the mould, Very Good design is about surfacing the most expected outcomes for any action which reduces friction and lets people get work done. And this generation of Generative tools is very good at identifying the most common/most expected response to a prompt.

It's a weird circle with these things. If you _can't_ do the task you are using the LLM for, you probably shouldn't.

But if you can do the task well enough to at least recognize likely-to-be-correct output, then you can get a lot done in less time than you would do it without their assistance.

Is that worth the second order effects we're seeing? I'm not convinced, but it's definitely changed the way we do work.

Not sure where you get 8 miles vs 11 miles (maybe the definition of a rural food desert?).

Low access is characterized by at least 500 people and/or 33 percent of the tract population residing more than 1 mile from a supermarket or large grocery in urban areas, and more than 10 miles in rural areas

(source: https://www.ers.usda.gov/webdocs/publications/45014/30940_er... )

Interestingly enough, this is measured by the euclidian distance, not by the actual number of miles required to travel.

One possible saving grace for Z is that, due to how expensive it is to keep around, video will probably disappear much more readily than text and photos.

I think that's the right broad question. Though LLMs properties mean that for some number of cases they will either make the results worse, or more confidently present wrong answers. This prompts the question: what do we mean by "quality" of results? Since the way current LLM interfaces tend to present results is quite different from traditional search.

If the current iteration of search engines are producing garbage results (due to an influx of garbage + SEO gaming their ranking systems) and LLMs are producing inaccurate results without any clear method proposed to correct them, why would combining the two systems not also produce garbage?

The problem I see with search is that the input is deeply hostile to what the consumers of search want. If the LLM's are particularly tuned to try and filter out that hostility, maybe I can see this going somewhere, but I suspect that just starts another arms race that the garbage producers are likely to win.

These policies are much clearer than they were when last I looked, which is good. On the other hand. Perplexity appeared to ignore robots.txt as part of a search-enhanced retrieval scheme, at least as recently as June of this year. The article title is pretty unkind, but the test they used pretty clearly shows what was going on.

https://www.wired.com/story/perplexity-is-a-bullshit-machine...

It takes this sort of critical scrutiny, otherwise mechanisms like robots.txt do get ignored, whether willfully or mistakenly.

There was a bunch of reporting on how AI companies and researchers were using tools that ignored robots.txt. It's a "polite request" that these companies had a strong incentive to ignore, so they did. That incentive is still there, so it is likely that some of them will continue to do so.

Anecdata, with a few exceptions, the VR games I tried were impressive as an experience, but not really all that fun once the novelty passed. The limitations of the format clash with the kinds of games that are being made so I often felt like the games were limited, or toy-like. I think the argument made by the article does hit on something about why VR isn't really getting accepted. The games are wrong: but they might be wrong because of the limitations vs expectations of the developers and audience.

The comfort issue is real too. Even with the fairly svelte PSVR2, it's annoying to wear those things.

This is, IMO, the better way to approach this problem. Minification applies rules to transform code, if we know the rules, we can reverse the process (but can't recover any lost information directly).

A nice, constrained, way to use a LLM here to enhance this solution is to ask it some variation of "what should this function be named?" and feed the output to a rename refactoring function.

You could do the same for variables, or be more holistic and ask it to rename variables and add comments (but risk the LLM changing what the code does).

Another way I've encountered this is performance vs results. Performance is the things you do that you believe will lead to results. Results aren't always in your control (especially in competitive environments), but performance absolutely is. It's a lot easier to feel you are getting somewhere when you focus on things that you control.

I dig this. Reminds me of column mode in MacOS's finder, which is similarly helpful in "rewinding" an exploration of a file system.

Would be interesting to rabbit other rabbit hole resources like Wikipedia or IMDB in this way too.

That's matches my experience as well. Really well put!

Yes, but if you don't have the LLM at the end, a good search (against a good corpus with the needed info) would still have given the user what they wanted. Which in this case, is a human vetted piece of relevant information. The LLM really only would be useful in this case for dressing up the result and that would actually reduce the trust in the result overall. Alternatively a LLM could play a role as part of the Natural language pipeline that drives the search, hidden from the user, and I feel that that's a much more interesting use of them.

The farther you go with RAGs, in my experience, the more they become an exercise in designing a good search engine, because garbage search results from the RAG stage always lead to garbage output from the LLM.

Seriously. This is why I don't like how much knowledge is being locked up in video platforms. I can get what I need to know from text so much faster than video unless the knowledge in question is inherently 3d/visual.

Four months for an offer sounds a lot more like a terrible salary negotiation tactic. The longer a candidate will wait for an offer, the less money they might be willing to accept.

It's definitely not going to get you the best candidate, but it might get you the cheapest one.

Re: Copenhagen + Amsterdam.

These cities are relatively flat, relatively cool (watch videos of cycling in Copenhagen and you'll see a lot of people in coats and hats), and you don't have to bike very far to get somewhere interesting. There's also the fact that there's a lot of bikes on the paths (which, for me, caused me to slow down considerably from my typical bike pace when I visited). This all combines to mean you aren't putting anymore effort in than a short walk: you just end up going a little farther in the same amount of time.

Re: the rest.

When I was a regular cycle commuter (in a Canadian city) I did pack a change of clothes, spare deodorant and/or wore a removable outer layer that resisted road gunk. Road gunk is _much worse_ on any route you share with cars. Dedicated bike paths tend to be quite clean. Paniers are far better than a backpack, since they sit on the bike frame, you barely notice the weight. But as other commenters noted: a little fitness goes a long way. After a year of commuting, my regular route wouldn't even cause me to break a sweat.

This, particularly the "do everything they say, even if you hate it" bit. Fighting a good teacher's process is just going to slow you down.

I was thinking something fairly similar. You could probably do pretty well with a basic NN setup this way, no need for an LLM. It wouldn't work on "never seen before cards" and would probably make some absurd picks when it's wrong, but I'd bet you could get to 90% accuracy.

Preserving author intent is definitely one of the major outstanding problems in collaborative editing. NN's have certainly shown the most promise in performing human-like judgement calls, but this would be particularly thorny: you'd need to them to resolve toward exactly the same judgement calls on data that may differ for periods of time (in order to keep the data consistency that CRDT's provide).

One advantage for me, is that systems like this let me think only about relationships. This is particularly good when designing, or exploring how to think about an existing system. So for me, I get into a flow state, just connecting the dots logically. Even better: I can adjust the diagram just by adjusting the relationships.

Miro and others, I am always thinking visually. It's another step removed from the real thought process.

As you point out, these systems sometimes produce difficult to read diagrams. For particularly important communications, I will re-draw these in draw.io, or a similar sort of tool. But then I am thinking more about presentation, rather than relationship design and the tool is better for the job.

Very neat! I really like the alignment tools, makes roughing things in quick.

FYI: There's an issue with opacity when you draw ellipses with a different color stroke and reduce the opacity. You start seeing the fill shape behind the stroked outline. There's probably a few ways to fix that, depending on how you implemented it. The easiest might be to shrink the inner shape depending on the stroke width.

Yeah, I would have been interested in this alternative hypothesis if it had any evidence behind it. Just a lot of "trust me bro" statements which don't cut it when peoples lives/livelihoods are in the mix.

I like the way you break down the emotional loop with criticism. For me, the way you've put this maps nicely to mindfulness practice. It's important to respond, rather than react to criticism. Taking a moment to listen to what your emotions are trying to tell you, and checking that against the facts of the situation is core to mindfulness.

I don't know how healthy this is but: I haven't run into many people who are harder on my work than I am.

I have an attitude that there's absolutely nothing I do that can't be improved. This may sound miserable on the surface, but for me it's actually quite freeing. It has the effect of making it easier to accept that things aren't actually going to be perfect (so it can help avoid the trap of over-engineering) and at the same time, it makes it much easier to have productive conversations with co-workers about what to improve/delete/rework because the existing product is something that could be better.

The other thing to think about, IMO, is that that PR comment, snarky or not, was something someone put the time in to come up with. For you. It's a gift. I've worked in places where getting any kind of PR comment beyond "approved" or "Fix your indenting" was an uphill battle. So getting a comment from a colleague that is meant to a) help make you better at what you do and b) help you both create something you can be proud of, is massively positive. It took time and effort for them to read what you did and think of a way to make it better.

Outside of FAANG experience.

Dental plans rarely cover anything more than basic maintenance costs. One procedure and you can end up pretty deep into pocket.

If you require regular medication, that can run pretty deep over a year too, since pharmacare plans have copays.

Yes, absolutely. There's a lot of people moving in to capitalize on the space between what people imagine they can do with LLM's, and what is currently possible. The difficulty in estimating what the LLM is likely to do with an output, the actual randomness that is involved in getting there, and the variability of the output creates a situation where you might as well put up an "INSERT GRIFT HERE" sign.

Even this post (which I think is making some pretty well informed and intentioned suggestions) exists largely because getting exactly what you want out of an LLM can be pretty difficult. Even fairly static tasks like data extraction can have aggravatingly variable outputs. I don't think that most of these are the _right_ way to get to the goal but are rather, largely clever hacks that can help a user try and nudge the LLM towards the desired latent space when adjusting the instructions fails.

Yes, this is true. But, I worry about how long it will take for the utility of "GPT-4" on my phone to be close enough to whatever is only possible through models running on large cloud platforms to make that choice relatively drawback free.

Is the curve of what this class of algorithms can provide sigmoid? If so, then yeah, eventually researchers should be able to democratize it sufficiently that the choice to use versions that can run on private hardware rational. But if the utility increases linearly or better over time/scale, the future will belong to whoever owns the biggest datacenters.

I give it about 5 minutes before the first child figures out how to jailbreak the tutor into helping it bypass the screen time controls preventing them from playing fortnite.

Claim verification is currently an open problem, as OpenAI admits to in their own promotional docs for ChatGPT. Issue number 1 here seems fairly thorny. And the second statement reveals the real problem: Even if the model was only trained on truths, would tuning it not to deviate from truth make it less able to answer questions?

ChatGPT sometimes writes plausible-sounding but incorrect or nonsensical answers. Fixing this issue is challenging, as: (1) during RL training, there’s currently no source of truth; (2) training the model to be more cautious causes it to decline questions that it can answer correctly; and (3) supervised training misleads the model because the ideal answer depends on what the model knows, rather than what the human demonstrator knows.

This is definitely not my area of expertise, but intuitively, it looks like increasing the complexity/varying the training techniques can increase the likelihood of correct answers, but I think the need to give the model leeway to let it work means that ultimately, either human or automated fact checking will need to be incorporated when using this kind of model for fact-finding questions.