cjf101

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.

She's just saying Ackman is throwing good money after bad. Don't see how that's arrogant, it's just a good analysis.

Based on his history, Sabatini will most likely end up costing Ackman time/money/trouble. There are other people that wouldn't come with that risk whose likely research output would be the same or better.

I mean, he's welcome to make the investments he wants, but it's a bad look.

Where I agree with you is that _policy_ is a much more direct and effective method of achieving similar ends. However I don't think eating less meat is comparable to carbon credits. The IPCC specifically lists it as a tried and true strategy for reducing greenhouse emissions.

The substitution of meat for other protein sources has real (albeit lagging) effects on the agricultural economy. Less people buying meat leads to less production of meat at scale, which in turn results in reduced deforestation (and potentially reforestation). Producing less meat leads to less greenhouse gas production at several points in the chain, most notably from reduced number of animals, but also during production and logistics. More forests contribute to carbon capture and biodiversity.

Probably the two biggest obstacles to consumer based solutions have been policy that subsidizes the cost of meat, and the cultural backlash to the idea of eating less meat (despite the many health + environmental incentives to reducing meat consumption overall).

I think this demonstrates how _insidious_ influencer marketing is. We feel like we're privvy to the behind the scenes decision making about this influencer or that one. But we aren't. We don't really know if they are really choosing the products they wish to promote, or if there are other considerations that influenced that decision.

The aim of influencer marketing is to leverage parasocial relationships to sell stuff, and it works because people think of influencers more like 'a somewhat distant friend' than a talking head or a traditional paid promoter.

I'm definitely a bit of a curmudgeon when it comes to email. Email is about communication and html gives people tools that they mostly use to make communication harder, not easier. For me, I prefer plain text email when I can get it. The main issue for me is that most of the uses of html email I see simply do not increase the utility of the email from the reader's perspective.

Large signatures don't communicate anything useful to me that isn't in the email header. The phone number can be nice, but you didn't need html for that. Your brand identity that takes up half a page (more since I keep email in a small window to the side of my monitor) does nothing other than obscure the content of the email. Trust me, sending me an intensely formatted email with lots of pictures does not increase the likelihood of me doing business with you.

99% of email I receive contains < 1 paragraph of useful information. I would prefer if people just sent that. Then I can read it quickly (since it uses the same font as everything else, not that faux handwriting font you found somewhere and thought looked great) and respond quickly and helpfully, which I hope was the purpose you sent the email in the first place.

There's probably a 1% where a nicely formatted document with graphs and images and headers are helpful. But you can just attach a document to the email if you're going to put that kind of effort it.

I strongly agree with the heart of what you are saying. One of the worst bugs you can have in a product is to have built the wrong product. A good channel of communication between people who understand the customer and people who are building and designing the product is not just valuable: it's critical.

There can be communication challenges between our disciplines though.

Frequent changes in direction can leave software in a bit of a shambles. This is probably the thing that creates the feeling that development is "a sacred space for creation". It takes time to both make the change, and end up with code that is pleasant to work with. On teams I've been on, this has sometimes led resistance to the marketing perspective.

I, personally, am often uncomfortable with the language of marketing (both on the analysis side, and on the promotion side). On the promotion side, marketing teams like to make claims that are aspirational or appeal to emotional needs, rather than actually true from a technologists perspective. On the analysis side, marketing often looks like it's working from incomplete information, and extrapolating trends that I don't feel confident exist. It feels like guessing, and I don't like to gamble.

These can be difficult gaps to bridge, but ultimately, I feel like we have the same goal: building the right thing for the right audience.

I wonder if videos are a little overused as tutorials. Videos aren't very searchable (yet), and they are unreliable to scan through (in most scrubbing implementations I've seen), so while they are nice to get started, they are hard to return to later to brush up on something.

I love this podcast + text approach in theory. A podcast to talk through the material, and a really well written page to reference. The CSS Grid page is a great example of this.

Aside: As someone who has been in web dev for a while, sometimes I feel like what I need is an "unlearn CSS" tutorial, to erase some old habits.

I'm not saying the writer's workplace has a meeting problem, but there's a thing that happens at orgs that do have a meeting problem. It becomes normal to double book people, because you can never find a timeslot where someone doesn't have a meeting.

It can also become known that you book focus time, and so people just ignore that you are busy, and book you in. It's also possible that once you are considered fairly respected that people trust you to decide what meetings you will attend.

IMO, Anyone with a calendar like the one in the article should start with a conversation with leaders at the company. This is massively bad for productivity if you doing anything remotely knowledge work related.

Booking focus blocks is a bit of a last resort. The problem isn't the calendar, it's the people and processes in the workplace.

Not the OP, but from my perspective, encryption is helpful, but a good portion of security is anonymity, and Signal requires that you use and leak personally identifiable information to even start using it.

It also informs you when people in your contact list are using Signal. It's probably not scanning through all of the phone numbers in Signal's database locally, so it is exfiltrating your contact list as well, exposing your network.

Personally, I'd prefer a model where I am not required to place even that much trust in the messaging provider.

There's also a lot less genetic information in a typical virus than in a human, so each mutation represents a much larger percentage change.

For context: The entire SARS-COV-2 genome is 29,811 nucleotides long. The human genome is around 3 billion base pairs.

It's really true that once you get probability, you see the world, read the news, and solve problems differently. If everyone had a good basis in stats, Vegas would be a Ghost town.

But if you spend time thinking about it, trig likely changed your perception of the world (and history) too. I think most mathematics do that to your brain, by virtue of expanding your internal language.

I can only imagine how crazy these conversations must get for the people actually trying to plan curricula.

I disagree with the impulse to discard calculus + trig.

You need more statistics than calculus to understand the current state of neural nets in Machine Learning, but you if you don't at least know the chain-rule, you're just using library code and not understanding even a basic perceptron.

And from a personal perspective: I've used so much trigonometry in my career (plenty of graphics work) that I'm not sure where I'd be without the time spent on trig and geometric reasoning in school.

I mean, I agree with their basic thesis that stats is important because of ML and the increasing practicality in using statistics and AI to solve problems. But I don't thin what they'd discard is well thought out in terms of what is going on. Nor is it meant to be, it's just a casual interview.

I don't believe it is using computer vision to read the pixels and translate that into data, no. It will still get the information about what's on its screen from the API.

So presumably it will get the same info you or I could get from the minimap (presence of units in a particular area but no details as to what) + detailed information about anything in its camera-like view. This will have some effect on how the bot manages its APM budget, since it now needs to consider camera management.

But honestly, from the perspective of the Alpha research, getting information from the pixels probably isn't an interesting problem. They can pretty safely assume that they could build a Computer Vision setup that could get the same info that the API gives, given enough time + GPU horsepower.

Any possible inaccuracy/delay that might result from using Computer Vision instead of the API can be simulated. You could add a response delay (already in AlphaStar), or a random chance that the API delivers incorrect information from time to time (which I don't think they do).

Thank you to you and your colleagues! I bounce back and forth between web and apple front end dev, and at first glance, this hits a lot of the highlights from both worlds.

Storyboard references made composing complex UI possible, but IMX there's almost always some breakpoint where it's just less complex to do it with code.

The funny thing is that while you are driving along a route, Apple Maps is actually pretty good at giving you exactly the info you need on cross streets, calling them out well ahead of time even when you don't need to make a turn there.

So to some extent Apple's designers and developers know the value of understand what streets are around, but that knowledge isn't evenly distributed/applied.

I think they are just using the Ctrl-Shift-Tab and Ctrl-Tab shortcuts to browse through tabs. Those work in Chrome and Safari, so you can keep your fingers on the keyboard.

I haven't found a way to use Contexts to introspect tabs (I suspect it would need some soft of browser extension to do so).

The Quick Tabs extension does a pretty good job if you don't mind invoking it after you get to Chrome.

It's not quite as elegant as finding windows, but it's pretty good.

From anywhere at all:

Contexts: Hold Fn, type "Chr" Chrome: Cmd-E Quick Tabs: Type tab name, Enter

I'm really similar. I hit a point with multiple monitors where I realized that the extra space was more of a distraction than a help. I now use a single monitor/full screen 90% of the time.

I use Contexts.app to search through all of my windows and so that task switching becomes more consistent/muscle memory-ish, which was really nice for staying in flow. I also use Spectacle so that I can easily window arrange for those times when side by side windows is useful.

And this is the relatively easy case of an app built using Apple technologies.

It can be much more difficult if you built a game using a cross platform game engine filled with binary 32bit libraries. If the engine is still around, you might be able to pay for a license for a newer version of the engine, but then you likely also need to deal with the SDK and format changes in the new engine as well as any iOS specific things you have.

Still, if software is your business and this software in particular makes you money, it's really not _that_ expensive to do.

Anything that can help make safer/more comfortable cycle routing is great step forward. Cycle routing is definitely more complex that driving because of the issue the article mentions.

I tried this out. My city is pretty notorious for having incomplete cycling infrastructure. i.e.) you get local areas that have cycle routes, but you often have to brave some craziness to get anywhere for real.

For my most common routes, it seems to be making the same choices as Google Maps. They definitely aren't the most optimal or comfortable choices, though I suspect this is at least partially a data problem. Steep grades, stairs, and difficult to find turns (more my city's fault that anything) are on the route, and it takes you though the straightforward but less safe connectors instead of detouring through neighbourhoods.

Perhaps one more tuning parameter is needed that would allow weighting stress vs distance?

Heh, that's awesome. I think it illustrates how confused people are in general about where bikes are supposed to be, which I think is a complicated cross of law, culture, cycling ability, and infrastructure.

The police have never hassled me for being on any road on a bike. But man, do the dudes in pickup trucks with truck nuts love to tell me to get off the road.

5 - 8. I was concerned at first. But there's so many of these on YouTube that have pretty high view counts, that my worry pretty quickly went from "my friend's kids are weird" to "kids are weird".

Have you seen the 2hour+ looping video phenomenon? My friends children introduced me to this. Their favourite videos are extended looping videos of the same 1-minute content. Songs, or funny moments, whatever. Their favourite YouTube activity is to see the same thing over and over and over for hours. They react nearly the same way to the same thing for the whole time. I'd never seen anything like it, but their recommender feed was full of them. I guess the algorithm works.

The article hints at this, but doesn't really point it out, when they talk about children living to see the same kind of thing repeatedly, to try and figure it out.

I'd be really curious to see if this actually helps the learning brain, or hijacks it, preventing deeper understanding through varied examples.

This is brilliant. DNS + API is a really smart way to build this. If you haven't done it already, it would be pretty straightforward to post a notification to the child's device (via a companion app on selected platforms) as an additional cue to let them know that they'd used up their time.

I'm halfway tempted to set this up for _myself_. You're absolutely right that 30 minutes is gone in instant when you start in on online entertainment.

From my experience you need three things:

0) You need a goal.

  * I want to give tech talks is good.  I want to talk about X at Y is better.
  * Keep it realistic.  When you achieve it, set the next goal.
  * Commit to the goal by signing up for it.  Runners do this and it works.  We sign up for a 5k race in a few months, and then we train for it, because now we're committed.

1) You really need to learn how to be a good speaker. This isn't actually particularly hard, but it does take a lot of work.

  * Find a good coach or program.
  * Give lots of speeches to small audiences.  People think large audiences are harder: they aren't.  
  * Get feedback from people you trust  Act on that feedback.  
  * Record yourself giving the speech and give yourself feedback.
  * Don't neglect the positive feedback.  Good public speaking is a lot about confidence, so learn what your strengths are, and bring them out.
  * Don't use slides until you can give great talks without them.  Slides distract you, and break your connection with the audience.  We think of them as speaking aides, and they can be, but they can just as easily do the opposite.
  * Learn to memorize your speeches.  If you're a musician you know why memorizing a piece is important. Same deal with speakings.
  * Murder your filler words.
  * Your talk can almost always be shorter.  The shorter talk is almost always better.

2) Have a tech topic you are passionate about. You should feel like you are the person that other people want to hear from about this topic.

  * Write about it for your self.
  * It's tech, so use it.  Build cool stuff.  Put it somewhere where other people can use it. Put where people that scare you can use it.  Maintain a project that directly relates to your passion.  
  * Develop best practices that you believe in.
  * Get involved with the global conversation about your topic.  Validate your ideas.

I live in Calgary. This thread is spot on.

I used to visit Edmonton often, and for the warm half of the year, it's just a more livable city than Calgary. Edmonton has actually walkable communities, a sense of genuine hospitality and a constant barrage of festivals. The Folk Music festival is a world class music festival (for the best experience, get on a volunteer crew, they treat you super well). There's plenty of other good music festivals though, like interstellar rodeo (not at all a Country Music festival), and Up+Downdown. Fringe is fantastic. Fort Edmonton is a fun outing if you dig history.

If you do locate up here though, get a group together and learn a couple of winter sports. Once I started snowboarding and snowshoeing, I really started to look forward to winter.

We have a flexible WFH system, in that we can work one or two days from home a week as desired/required.

It's a 20-25 minute commute to work, and I really notice the difference in the quality of my day when I work from home. I can work, go to the gym, make food, do chores, and still have time to do something worthwhile after work hours.

When I commute, all of the little things end up and my day feels so much shorter.

That said, I like going into the office, even if it is cubicle land. Spontaneous design conversations are valuable. If people are having a problem, I can hear it, and jump in if I can contribute. It would be challenging to be 100% remote in my role, which involves communication, mentoring, and team organization in additional to technical work.

If I lived two blocks from the office, I'd have the best of both worlds :)

A number of my colleagues print to paper for code review. They find it easier to read and markup in that format.

I personally can't stand the idea of using that much paper, but I dig the future-y idea of printing to a read/markup slate, and automatically posting the markup to phabricator.

I imagine legal and other business functions that are review heavy could similarly benefit.