I think there’s a bit of parroting going around but LLMs are predictive and there’s a lot you can inuit a lot about how they behave just on that fact alone. Sure, calling it “token” prediction is oversimplifying things, but stating that, by their nature, LLMs are guessing at the next most likely thing in the scenario (next data structure needing to be coded up, next step in a process, next concept to cover in a paragraph, etc.) is a very useful mental model.
I would challenge the utility of this mental model as again they're not simply tracing a "most likely" path unless your sampling methods are trivially greedy. I don't know of a better way to model it, and I promise I'm not trying to be anal here
Agreed - I picked certain words to be intentionally ambiguous eg “most likely” since it provides an effective intuitive grasp of what’s going on, even if it’s more complicated than that.
Honestly, I think the best way to reason about LLM behavior is to abandon any sort of white-box mental model (where you start from things you “know” about their internal mechanisms). Treat them as a black box, observe their behavior in many situations and over a long period of time, draw conclusions from the patterns you observe and test if your conclusions have predictive weight.
Of course, if someone is predisposed to incuriosity about LLMs and refuses to use them, they won’t be able to participate in that approach. However I don’t think there’s an alternative.
This is precisely what I recommend to people starting out with LLMs: do not start with the architecture, start with their behavior - use them for a while as a black box and then circle back and learn about transformers and cross entropy loss functions and whatever. Bottom-up approaches to learning work well in other areas of computing, but not this - there is nothing in the architecture to suggest the emergent behavior that we see.
This is more or less how I came to the mental model I have that I refer to above. It helps me tremendously in knowing what to expect from every model I’ve used.
The screens included in these kits are glossy, however the board can support pretty much any screen up to 13”, including matte and flexible. On the ghosting - we’ve been thinking a lot about Wayland drivers and other software solutions. We’re working on an SDK that allows a compositor or other software to not just refresh portions of the screen but also change the display mode for certain portions as well.
This has been tried so many times[0][1]… heck even I made an app like this once using WiFi direct.
The idea is so solid and yet there are just enough pitfalls between Bluetooth reliability, platform differences, getting critical mass for effective relaying…it’s such a bummer that we can’t figure this one out. Decentralized message relays have the potential to work anywhere, be fully private, extremely difficult to block/censor, and (in theory) can scale indefinitely.
I worked with BLE recently and found it very reliable. Also I was surprised at range it was able to communicate. Like 10 meters with walls - it works. Similar to WiFi.
I think, it's all matter of inventing a proper protocol on top of it, and enough work-hours put into the implementation to make software reliable.
The FindMy network is already doing this at scale with location data. Offline messaging might be too niche of a use case for Apple but surely they’ve considered it.
It very much doesn't. The FindMy network just scans for BT beacons around, and reports what it sees to Apple's servers via the cellular network. It very much doesn't do mesh networking at all.
The missing piece might be thinking of these mesh networks as a new "tier 4" network and routing to the larger internet as quickly as possible.
One huge advantage of this (beyond better networking) would be that apps could use existing IP APIs which would help abstract away vendor/implementation specific problems and improve adoption.
Note that this doesn't necessarily depend on a tier 3 network so you could still accomplish your goals. The internet already has enough features to support partitions and local discovery.
You of course still have to have fixed administrative domains but in reality you always do. Someone takes initiative and sets up the group chat/gets their friends into it. I think if you're willing to mentally separate the network topology from the administrative topology this could be solved.
Of course that's really boring. It's just MDNS and ad-hoc Wifi plus some routing. Everything is pretty much already there (although iOS probably won't let you do it, as usual.)
But does WiFi Direct work on iOS so we can be on a music festival and mesh between your friends (Apple) mobiles? This would be much better than Bluetooth. I read about the MultipeerConnectivity framework but I am confused because in a recent discussion someone says that WiFi Direct has (development?) limitations? ELI5 please.
Apple uses AWDL, a proprietary protocol that predates WiFi direct. WiFi direct is kind of a mess, and then there’s WiFi aware. Last I heard when I was working in this space a couple of years ago there was supposed to be a WiFi direct 2 that would (as usual) unify standards and solve all the issues.
Which is hilarious because WiFi Direct is a mess. Possibly this will be the kick the industry needs to get a decent P2P ad hoc wifi protocol going, but more likely Apple will just implement what the EU demands and it will be just as terrible as wifi direct is today, so nobody will use it anyway and AWDL, as bad as it is, will remain the usable protocol in this space.
I’m not sure it would since the whole Bluetooth stack is already designed for P2P. Apple’s network is a true mesh though and not point to point. If you share locations with your friends on FindMy using iPhones or AirTags you will get updates via the encrypted mesh network (that includes other peoples’ iPhones) without being connected to a WiFi or cell network.
My interpretation here is that the location of a missing device can be passed around via Bluetooth mesh but that the last send must be sent via cellular or WiFi back to iCloud. So it’s P2P to relay getting location info back to final device with internet connection back to iCloud but the finder still need to connect to iCloud to view updated location.
“Find My works offline by sending out short range Bluetooth signals from the missing device that can be detected by other Apple devices in use nearby. Those nearby devices then relay the detected location of the missing device to iCloud so users can locate it in the Find My app—all while protecting the privacy and security of all the users involved.”
I really wish a "major" messaging service would adopt something like this, at least as a fallback. Don't give me two apps, give me one. This seems right up the alley for a project like Signal. Like a way to wade yourself into decentralization without necessarily needing to go all the way. Hell, it could save bandwidth just by file transfers alone and certainly it is up there with the mission of privacy.
I think it really will take a bigger platform to make this possible because you need an already existing network. I doubt Apple would ever do it, but hey, I mean text messaging and calls through Airdrop? Pitch it as for emergencies like when the phone lines go down? These are legitimate use cases.
I agree with the general sentiment but I hesitate to rely on any platform being around for a long time. It would be cool if they could interop with other platforms supporting it, though.
Fantastic work. This is great example of how good execution is what really matters - not just good ideas. I’m sure I’m not the only one who had an idea similar to this at some point - mine was called “nanowriter” and was meant for NanoWriMo (RIP)[0] but I lacked the coding ability and executive function to actually make it.l at the time. Enso is gorgeous and… exists, and therefore is infinitely better.
Lottie is quickly becoming the de-facto standard for UI animations, but live, long running vector animations aren’t really something I’ve seen much of - at this point, video compresses well enough that people will simply use an mp4 or webp for that use case. I know that’s not what you’re looking for but since it isn’t a common use case, I haven’t seen much support for it. Lottie is perfectly capable for this use case, provided you don’t want audio.
I’m seeing some debate on Deno’s decision to ensure Node compatibility, apparently as it gives up a core value prop of early Deno to try and hit the reset button.
Can someone help me understand what was lost here? Is there no longer a way to use Deno without using the Node ecosystem?
It adds complexity. Instead of being forced to do things the Deno way, now users don't migrate at all from node code and then (not unreasonably, due to their mission of node compat) complain when they don't work. A large amount of feature work stalled due to the node compat work for over a year. Even now, in 2.3.0 over 30% of all listed changelog items are node compat. They started pushing you to use config files to better integrate with your dependencies as a new default. etc
Does anyone remember the “Put a bird on it!” Portlandia sketch? As if putting a cute little bird on something suddenly made it better… my personal running gag with SaaS these days is “Put AI on it!”
This is the trap I fall into. I have had so many amazing colleagues and I want to do right by them. Sometimes it’s been trench camaraderie, sometimes just really great working relationships, but I almost always feel like I owe it to my fellow employees to work hard, do well for the company, etc.
It’s taken me a long time to learn, but that form of loyalty doesn’t equate to employer loyalty.
I’m working on a live voice chat for settling rule disputes during tabletop gaming. It uses OpenAI’s realtime API to allow anyone to call in a “ref”, trained on the rules of the games, to answer questions about the rules.
I have four books that I got from my dad - one for construction and repairs, one for electrical, one for plumbing, and one for landscaping. They’re in simple English with excellent illustrations. Using those books, I’ve been able to fix/improve pretty much everything in my house for ten years.
Could you list the titles/author of these books? I'd actually really like to get more into DIY home improvement, and having a starting reference like this would be really valuable
The electrical and plumbing books are from a series from Home Depot called "1-2-3": "Plumbing 1-2-3" and "Wiring 1-2-3". I love them but some other posters have shared books that look even better; I plan to check them out. The construction book is "Modern Carpentry" by Willis H. Wagner. I can't find the landscaping book - haven't needed that one as much - but in my stack I noticed Reader's Digest's "Home Improvements Manual" which I've also used quite a bit.
It's still a human doing it and not an AI or robot. All of these jobs are not uniform enough for a robot or AI to be able to handle everyone's uniquely built house.
you also got standardized parts and plug-in systems that use them; standardized tools that were available for a reasonable cost; and you have basic health and mobility
