The whole "let's make an Android device for seniors/the blind/whoever and pin it on an ancient Android version/spec it really low but charge a premium" thing is unfortunately not new.
If you look at the Blindshell 2 and 3 you see plenty of this happening as well, not to mention way more expensive hardware like the BrailleSense Polaris which last time I checked ran on Android 5, scary "Mind updating right now please?" Google Play warnings and everything.
If it helps any, Talkback does have a set of keyboard commands it responds to, perhaps making your keypad buttons fire those means the implementation might be less brittle but given what you've already had to deal with maybe just not touching it ever again might be thebest thing at this point :)
Disclaimer: Actually blind Android user/developer
and then there's the abstraction of 'activation' used in neuralnets that is very useful computationally, but omits a lot of the complexity in actual neural signals. you don't get fine-detail in the timing of spikes or the fact that neurons use different types of neurotransmitters.
A lot of people seem to think that training is limited to the 10 bits/s of processing mentioned here. But raw sensor input (mostly visual) is in the order of 21 quadrillion bits per year [2], according to the same paper. It seems likely to me that learning uses much more data than is ultimately expressed in behaviour.
[1] https://www.sciencedirect.com/science/article/pii/S089662732...
[2] Assuming 16 waking hours, and 10^9 bits/s of input that the paper mentions. 16*3600*365*10^9 = 21.024*10^15
I don’t think that’s a fair comparison since (as far as I know) humans aren’t born with randomly initialised neural weights that are setup in a very generic structure, but have at the very least extremely strong inductive biases. I think it’s fair to count the hundreds of millions of years of evolution between the first animals with a primitive neural system and humans today, and the amount of “tokens” that were “trained on” is likely enormous. I think it’s better to consider the 200M tokens as a sort of “fine tuning” to a sort of genetic “pre-training”.
You lose some character moments, but you get to everything the parent post described far quicker. And it’s well, well worth it, especially for those of us working with AI.
PoI asks these incredible questions: how do you design an AGI system that can optimize for saving lives, without it taking a strictly utilitarian view that leads inexorably to it trying to take over the world?
And how do you, as a human trying to get the best of both worlds, keep yourself sane when the tradeoffs are visceral and painful and demand a sacrifice?
You get to know a group of lovable, flawed, but highly competent weirdos who are trying to muddle through this. (With Amy Acker’s incredible performance a particular highlight!)
(It’s also the source of the four-people-pointing-guns-at-each-other-in-church meme, which is delightfully lampshaded in universe.)
Highly recommend.
I like to compare how similar or different The Machine is to current state-of-the-art LLMs. Interestingly, in this fiction, The Machine's preferred way of communicating doesn't seem to be through text, or even artificial speech. It can see almost everything, and seems capable of improving itself, but it feels "intelligent" in a way that humans or LLMs aren't. Maybe the writers thought that giving it the ability to write text or speak directly would make it "less" in many ways (less mysterious, less machine-like, less intelligent, etc.). It also gives the impression that you cannot really guess what it "thinks" or fully understand it.
I'm also realizing that most of the time, characters don't really have a "conversation" with The Machine, whereas it is a core feature of current-day LLMs. And The Machine doesn't even "create" its communication channels, it uses existing ones (phone booth, radio ...).
In case you’re looking for some more shows related to AI, Cupertino by Robert King and Michelle King is currently airing and is pretty entertaining.
One of my most fav TV Serials to have come out of the West.
It's really hard for me to believe American kids suddenly can't get enough books with male protagonists. If that's the case, the fault must lie with whoever's choosing books for those kids.
When you look at where money is spent and by whom, keep it in mind as you shift the demographics of who's watching / reading. The money moves with the groups interest, and it doesn't necessarily get filled by the new demographic.
I didn't go into management because I didn't think circumstances were stable enough. I did see four co-workers go up the ladder, with disastrous results. Two of them I tried to warn, the other two would have been a waste of time to warn. By the time I retired, two had been fired, one drank himself into a coma, then passed away. The fourth is still working, but became so hated, the company keeps moving him to different posts. All four of these cases were promoted by management personnel who were "people persons", but were terrible decision makers.
There's an old saying, "Ain't nothing to it, if you're born to it." I know it's too simplistic, but it's based on a lot of reality.
Both is fine, these are just different jobs. However, the problem is that the manager is 'above' the engineer in the generally accepted hierarchy. Consequently, engineers are pressured to become managers for the development. As long as this hierarchy exists, the problem will exist.
Organizations where they get away with this are miserable to work for, and their influence ensures nothing merit-related ever happens again. Literally the source of the famous expression “A-people hire A-people, B-people hire C-people”.
There is no going back to writing code the old way. I don't see that happening in the industry, ever.
The grievances listed here are 100% real, but not fundamental - with some practice you can prompt an LLM to great effect without it taking the reins from you. If you let it write the code, it does become a management job more than an engineering job. That said, I do believe that there must be a way to structure your workflow such that the LLM becomes your read-only pair programmer/rubber duck while you yourself get to do the fun parts of coding.
Yet I find myself using AI coding agents more and more. In this day and age, your choice is to either use them, or be stomped into the ground by programmers who do.
AI still can't write code at my level of quality, but it can do so at adequate quality now; and sadly, that is "good enough" for the real world.
I haven't touched a line of code in over 6 months, didn't even look at it, but the work I do is faster and higher quality than ever.
Not sure how I feel about it
AI does benefit from learning to use skills, and being instructed with good structures (like “save [specific important kind of detail] in memories” and “plan before implementing anything”). But none of these memory solutions seem better than any other, they’re basically just different folders. Your CLAUDE.md can say “see AGENTS.md” and AGENTS.md can say “write memories to .agent/memories/“, and I’d be interested in a measurement, but I assume it performs almost the exact same as native CLAUDE.md and .claude/memories.
One thing I’ve noticed AI does much better than me, is not have OCD over the “perfect” solution.
I usually read the comments before deciding if the article is worth my time.
(Integer/parseInt my-variable); this calls Integer.parseInt(my_variable) in Java-land
I tried to find examples for how to call Rust code and the best I could find is:
https://github.com/csm/clojurust/blob/b847a29b46e2b6b3fbe624...
It looks like you need to do a lot of leg work to introduce Rust functions into Clojure namespaces. I guess there is no way around it and this is not a criticism of clj.rs. But it doesn't seem to be as seamless as on the JVM.
This would sell
Now we are in a situation that anyone can do anything with any data. And this was not what we were thinking. There needs to be a legal reset.
A new judicial ruling that forbids such permissions as its terms of service. That calls such permissions exploitation. That labels them as a force of evil (like the patents meant to favor edison’s movie cartels) and invalidates all powers they hold. And a judicial ruling that unilaterally limits all prior TOS agreements to be for just specific existing uses. And all future TOS’s to be opt-in, versus being forced to accept or withhold service.
If I subscribe to Google, and Google changes its TOS, why the hell should I accept a new TOS just to get access to my data? Big large govt agencies negotiate TOS’s and hold the companies to those outdated TOS’s so the companies understand well that they cant just unilaterally change it and must maintain older processes that respect legacy TOS’s.
It's a shame this book by my MSc dissertation supervisor isn't coming out until next June, because I'm pretty sure there's a bunch of stuff in there about this kind of neurobollocks: https://www.penguin.co.uk/books/479845/neurobollocks-by-devl...
Maybe it would be better to measure uptime not as what percentage of time the site is up, but instead as how likely a randomly selected visitor is to be able to access the site.
That would adjust for the fact that the window of peak demand, e.g. the US workday, which is when Github tends to go down, is also when people need the service the most.
GitHub has outages all the time without that page showing anything.
Another comment pointed to a third party alternative which is more accurate – https://mrshu.github.io/github-statuses/ – but that alternative isn't showing an outage right now either, since I believe it is being fed by GitHub.com's incident data.
I usually check this and then doubt my network might be at fault if they don't show outages.
I could, before LLM-authored PRs. They broke all my heuristics for what to check. LLMs can write code that looks incredibly professional and well thought out, and nowadays they almost never make the usual human mistakes (like off-by-one errors). But at the same time, they can make incredibly silly mistakes of the kind people would never make (like inventing brand-new parameters for some remote API call).
It feels like everything I’ve learned over the years about reviewing code no longer applies.
Anyone else experience this?
On topic, regarding these changes. I'm not sure what all changed, but I like having the status checks and merge button at the top of the page now.
At least for me. Am I missing something?
I wonder, do they not notice? Do they not think we do? Do they not care about the quality of their copy? If not, why on earth would anyone conclude that they care more about the product? There is so much evidence that they do not care, this is yet another piece of data to add to the pile.
Again, it’s just so sad that a company which used to be justifiably beloved has become so miserable and incompetent.
Code review is great if the code isn't a surprise, the architecture was already agreed on, the AI reviewers are already happy, and you just need a final pass.
Code review was already terrible before AI if you were slopping down weeks of work on a poor teammate to sift through.
I tell the younger engineers at my company the same thing today I said 5 and 10 years ago. Make the review easy for another engineer to do it, and they'll do it. Make it hard on them, and you won't ship shit.
Their support has acknowledged the issue to us over a month ago but no fix as of now.
Ugh
There’s SO much to be done around Code Review, I’m very surprised GitHub isn’t doing more.
They have the reach already, they have years of real review data. And yet, the PR itself is not using any of the newly available methods for more efficient review.
Agents are hammering PRs out left and right, and yet here I am still staring at the raw diff of 35 PRs and end up being called the bottleneck.
- Dual Pentium-II CPUs
- 512 MB RAM
- 100 Mbps wired ethernet
- Red Hat Linux
Mass storage was larger, but not by much, and the flash eMMC is probably faster than the spinning hard disks we used back then.
Possibly of note from my services list:
victoria-metrics : currently at 230 MB ram because I send it all kinds of metrics
victoria-logs : 114 MB ram -- log aggregator
vl-agent : 25 MB ram (log forwarder for above)
ntp : 2 MB ram
librespeed : 16 MB ram (LAN speed testing, useful for wifi / quick ethernet checks)
corentinth/it-tools : 5 MB ram
mqtt mosquitto : 5 MB ram (message broker)
smokeping : 52 MB ram (latency history)
https://github.com/apenwarr/blip : 2 MB (live DNS latency checker)
lighttpd : 2 MB (static web server serving a single homepage)
Happy to answer any question and take suggestions as to what other software would be a good addition!
For anyone who can load APKs on their TV and wants a more integrated solution, this alternative client works wonderfully https://github.com/yuliskov/smarttube
It's at the level of polish where people didn't even realize it wasn't the official client for quite a while, until I turned on SponsorBlock that is.
Is it just me or does anyone else feel that 200 MB is a lot for a native chat app written in Rust, that's essentially a remote interface for the WhatsApp on your phone?
I remember back in the day, apps like Miranda IM (and early versions of Trillian) used only digits of RAM, even when you had multiple plugins talking several protocols at the same time, whilst having a ton of advanced features not found in WhatsApp. Yes I get that times have changed and software has become bloated etc, but 200 MB still feels a bit too much.
it's an incredible waste of resources, especially at high resolutions, to redraw a Gui application like you would a videogame at 60 FPS. These immediate mode UI frameworks were made in particular for that case, to be embedded in real time software, not for a UI where the state is idle or changes only occasionally.
Egui is definitely not what you want to use on a 4K screen. It is only good for tiny screens of embedded devices.
Why there is no reactive UI that tracks all dependencies and redraws only what needs to be redrawn? (not React, it also re-calculates (but not redraws) everything on every change).
Also why does it visually look like bad version of Telegram? And Telegram can run without a phone present.
> Using an unofficial client may be against WhatsApp's terms of service and could get an account suspended.
Same for all of these unofficial clients.
https://en.wikipedia.org/wiki/Effects_of_nuclear_explosions
https://en.wikipedia.org/wiki/Effects_of_nuclear_explosions_...
These claims are more about politics than anything. Its an attempt to retcon history to vilify the US. The Japanese had never been invaded and that was a huge point of pride. They were training the civilians for a guerrilla war. Just because you can find a person in the Japanese government who wanted to surrender at that point (and there were) doesn't mean that's the decision that would have been taken.
PS That 3rd point is especially specious...if it were true, no 2nd bomb would have been used. PPS Stuff like this is why historians don't like nor will engage with hypotheticals or the alternative history folks. Its clearly politically motivated, not a good faith attempt to understand history.
In contrast, the first allied nuclear reactor was a literal pile of uranium and graphite on a squash court.
[0] The Bastard Brigade (great book on all of this)
Every time I re-read it, it makes me want to take up nordic skiing. It is absolutely up there on my list of greatest true stories ever told with Touching the Void, The Emerald Mile, and a select few others.
https://www.abebooks.com/Winter-Fortress-Epic-Mission-Sabota...
" Fermi writes that "The results of this experiment was [sic] somewhat discouraging" presumably because of the absorption of neutrons by some unknown impurity. So, in December 1940 Fermi and Szilard met with Herbert G. MacPherson and V. C. Hamister at National Carbon to discuss the possible existence of impurities in graphite. During this conversation it became clear that minute quantities of boron impurities were the source of the problem. "
Furthermore, if the GPL is really as unenforceable as you say, then there’s no reason why their license change is “punitive”. Either the GPL works, in which case the Asahi Linux maintainers have the right to restrict use of their software, or it doesn’t and there’s no reason why anyone should be upset by a gesture of free speech.
The Asahi Linux maintainers have the opinion that LLMs are harming their ecosystem, and if they want to prevent people from using LLMs on their software, then they should be able to do so without being unfairly attacked.
> Luckily for us, Sven was up to the task of dealing with all this nonsense. It just took slightly longer than expected! While the patchset to enable Thunderbolt was sent upstream first (another massive win for us!), we have also pulled it into our downstream kernel and are now shipping it to users!
> But there is a catch. Currently, only USB3, XDomain and Thunderbolt Networking are supported. You can plug in USB devices or connect two Thunderbolt computers together and make them ping each other, but that’s about it. PCIe and DisplayPort tunnelling — required for most Thunderbolt docks — are not supported. Yet. Watch this space.
> An enormous amount of gratitude is owed to Sven, who has stuck with this thankless chore for nearly 5 years in between all of his other incredible work.
Thank you, Sven!
This is a wonderful thing to read. I've been a proud financial backer of Asahi since the first announcement of their Patreon account (now OpenCollective). I've only installed and booted it once on an M1 MacBook Air to kick the tires. I liked what I saw. It worked well enough to validate the work being done. I then reinstalled MacOS so that my partner had dedicated modern laptop. I also doubled my donation.
> Some time ago, we were made aware of multiple attempts to exploit our permissive licensing to hijack our work, with a direct view to replacing the project and aggressively monetising Linux on Apple Silicon. In some extreme cases, in-development feature code was being lifted straight out of our personal forks and published without attribution or lazily rewritten with LLM assistance to obfuscate its origin. The MIT licence explicitly allows this, and has no carveouts to protect against bad faith actors.
> As such, we made the difficult decision to begin the process of transitioning all of our repositories to the GNU General Public License version 2 or later. Going forward all Asahi Linux code will be licenced under strong copyleft licences only. Those who do not learn from history are doomed to repeat it.
Oh man, I just doubled my donation again. I'm now at 4x my original donation based on reading this because hell yeah. Congrats to the team doing this stellar work. Now that my partner has a MacBook Neo, I'm gonna wipe that MacBook Air and test out the latest release. I'm so happy folks are doing this.
I support Asahi due to ideological conviction and self-interest (in case I ever want to have an alternative on Apple hardware) rather than because I want to replace MacOS. But who knows. If Apple's rent-seeking behavior with the App Store grows and further disrupts my opinion of them, maybe I'll end up on Asahi full time.
There are a few significant reasons worth considering when getting down to the level of OS and hardware drivers. Long term maintainability is one. But another big one is that such software is capable of physically and permanently destroying the hardware it's operating on, with the textbook example for Macs being the speakers found in Macbooks. The only way to get the kind of performance they offer out of that physically light/thin package is to give them the power to destroy themselves if run at full force continuously but then run it from software so that they aren't run at full force continuously, but rather take advantage of the headroom that offers for very brief audio spikes whose overall integrated power still falls below safe limits. This all needs to be done from software since response of real world analog isn't instant. That enables much more range then if they were hardware capped at the maximum level they could run constantly, but it also means that if they're driven wrong, poof. The drivers have to never go to certain bad states, purely from software.
LLMs can't offer those kinds of guarantees, they don't really "know" how things work and then produce something deterministic off of that. It's even worse if, as-is the case with this sort of vibecoding effort, the LLM operator is in fact depending on the LLM for that part because they don't know how it really works either (vs someone who is just using it for boiler plate and knows exactly where the critical paths are).
Obviously everyone is welcome to roll the dice with their own hardware as they wish, and it may be ok. Or if using really old hardware one might just not care. And not all hardware is as sensitive as modern notebooks of course. Just be aware the analog world can be messy and you might not actually have the guardrails to screw up that you're used to in a higher level software environment.
>Otherwise one can also say Asahi is compiler-generated slop compared to hand written assembler code.
No, one cannot say that. Or at least not without looking really stupid.
Unless your meaning is that you already have work-in-progress code that you are working on to make much better whereas they have not even started, I don't see how your point makes sense.
But even then it debatable whether making bad code good is better than creating good code from the start.
IIRC Asahi has finally acquired Thunderbolt support but that support still doesn't have the pieces needed for external displays.
Some of us still care.
This is confusing - your website shows M4 as Wip (i.e. pre-alpha) and only M4 mini as alpha. M5 is not listed at all.
Wip is described as "Bring-up in progress. Not usable yet." So is it "not usable" or "all of M4 (...) working"?
There's a reason no Omarchy patch of any kind has made it to the kernel list yet. I did personally look into some of their code and it is certified, hacked-togeter slop.
Asahi at least guarantees the quality of their stuff because they upstream everything, sooner or later. Gravity apparently also intends to be upstream-first.
I am just surprised that other features like support of the newer CPUs still get a higher priority than this one. I am not complaining or anything, I am not even related to the project in any way, so not in the right place to ask for anything, but I really am surprised that this seems to not to bother most of the users?
But most people just want stuff to work, so it's odd to complain about it when others do manage to get it to work using AI to build on top of their prior permissively-licensed work.
Gravity Linux vs Asahi Linux
Gravity Linux and Asahi Linux are two independent projects working towards the shared goal of a strong, open OS ecosystem on Apple Silicon devices. Gravity Linux was initially forked from Asahi Linux over differences in policy, namely regarding LLM use. We have tremendous respect for the Asahi Linux project and emphasize our shared goal of running Linux on Apple Silicon.
I am not one of those people, as I do not do cars, but I would hazard a guess that most car people over a certain age do in fact feel very strongly about what modern cars have become. The ones who grew up without knowing what ownership feels like less so, which reflects the exact same dynamic with younger generations who are growing up without ever knowing what it feels like to own the hardware they use.
I agree it was a poor choice of license and that GPL helps build a better community, I'm glad they mad ethe license choice
“But software which OpenBSD uses and redistributes must be free to all (be they people or companies), for any purpose they wish to use it, including modification, use, peeing on, or even integration into baby mulching machines or atomic bombs to be dropped on Australia.”
—Theo de Raadt (2001)
edit: worth noting this example, where DigitalOcean started cutting OSS support only to pledge millions of dollars to Omarchy. No wonder people consider the project a leech to the OSS community, but also consider correlation != causation: https://itsfoss.com/news/digitalocean-open-source-credits-en...
- MacBooks are still among the nicest if not the nicest laptops hardware-wise, so some people like to run Linux on it.
- People have older Macs that they like to repurpose. E.g., we still have an M1 Mini that was used by our daughter before she got handed down a MacBook from my wife. It's now collecting dust, but I might install Linux to use it as a headless server.
Also, this is Hacker News.
It seems like many times it's open source but "just to look".
Is this referring to Omarchy? Anyone have context? TBH I'm surprised their work was ever MIT-licensed instead of GPL.
Spoiler: We have Vulkan 1.4 working, all of M4 and most of M5 working, are making significant progress towards full suspend to ram (no version of Asahi supports this) and this will all be shipped to users very soon!
3=(sin(arctan(sin(arctan(sin(arctan(2))))))^−2 − (2)^−2
I've been using Linux on the desktop from 2000 onwards, and exclusively about 2005 or so. I've been a law student using free software, and I wrote my lawyer's thesis in LaTeX.
I practice law using LibreOffice, and I'm having no problem using it. I keep a Word365 install in a Windows partition for compatibility purposes, but I almost don't use it (the fact that it forces that hideous ribbon interface on me doesn't help... Hey MS, give us a proper WIMP menu back!).
For PDFs I use Okular, which handles digital signatures well and also offers a bunch of annotation features.
My Linux setup is Slackware-current, right now I'm using Plasma 6.8.
If anything, Linux and free software helped me work more efficiently.
At the end of the day, every firm needs some form of IT. Above a certain scale you are hiring IT help to manage it anyway, might as well use free software where you can over MS. It’s not as if managing MS for a large org is particularly easy.
It's definitely not categorically malpractice to not use Word. Depends entirely on the situation. If a court demands docx and you file an rtf, or you refuse to read a client's redlines because they're in Word, sure, that could be malpractice.
The only proprietary software that I think _might_ be malpractice not to use is a legal research service with a capable citator, like Westlaw or LexisNexis. If you rely on a case that was overruled, you can't defend a malpractice suit by saying, "I didn't want to pay for the software."
I hear CourtListener (a free legal research tool) is working on a citator--hopefully it's competitive with the paid services.
Like Intuit raising quickbook prices by 20% every few months.
Her high school took delivery of its first ever computer, a PET with a chiclet keyboard. She was assigned to take it home over the summer to learn, and figure out how to use in the curriculum.
A few memories- it was the first piece of electronic equipment in the house with a grounded plug, and the house had no grounded outlets. Most people would have used a cheezy adapter, but Dad sledgehammered an 8’ copper rod in the backyard and installed a single grounded outlet for it.
The cassette loading could be very fiddly and slow. A favorite game was “Hangman”, played that for hours. Observe the beautiful graphics.
https://youtu.be/jnj2ZgOYYDQ?is=zZauKOy5f95wDqrg
The PET went back to the school in the fall, and I started begging for a PC of my own. For the longest time I was constantly asking for a Timex/Sinclair. Christmas 1983, a TI-99/4a appeared under the tree, thanks to them being sold far below cost after failing in the market. That lead to an Apple //e in 1987, and a Dell 486 DX2-66 in 1993.
Which parameter to vary depends on the result of the test, but the MaxxBench route of varying the task spec seems like a good default, though not always the most informative.
In one of my private benchmarks I realised many of the less capable models didn't fail because they submitted the wrong answer, but because they exceeded their budget for number of tool calls they were allowed. The more capable models never got close to the tool call limit because they tended toward combining tool calls (e.g. chaining multiple shell commands in one Bash tool call).
Was my benchmark actually measuring ability to combine tool calls rather than reasoning ability on the problem at hand? Well, easy enough to find out: re-run under a different tool call budget and see how much the scores change.
(Next problem: is ability to combine tool calls a useful thing to measure, i.e. should it confer an advantage in the benchmark? It could be if it is a sign of planning multiple steps ahead.)
It's not unreasonable to assume Kimi 3.1 is also a distilled model even though it's large (the difference being that it's a distilled from US models). I wouldn't be surprised if it also collapsed on this benchmark.
The practical takeaway, however, seems to be an interesting "don't use max mode". The passage about deepseeks weird API behavior dropping to pretty much zero the moment you used high says a lot.
Also, personally, even though Opus is clearly so much ahead of the Open Source models after the changes, I think the adjusted score is still very much competitive. I see other commenters using this as proof to justify why their opinions of the open source models - but a drop from 74% to ~50% is, while a big drop, still about a ~50% pass rate afterwards. That's far from useless for coding, especially since DeepSWE seems to be such a difficult benchmark. It simply seems to more accurately reflect the capabilities of these models.
All in all, I'd love to see more of such articles, and will probably keep an eye on this one - if they continue "benchmarking" these models this way.
Edit: I think that's the answer. This page lists the score as 53.3% which is close to the 54.4% reported here https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731 for the 0731 update. The original from April was called preview, doesn't list the score, but in the comparison with the update it scores 7.3%.
Look at GPT 5.6-Sol vs. GPT 5.6-Luna in particular.
(The GLM pair shouldn't be expected follow this pattern, because they're misnamed: GLM-5.3 and GLM-5.3-flash are distinct lineages. The -flash is an early release of a new pretrain on a new architecture).
If this *is* the main factor, I'd expect Kimi K3.1 and Gemini Carbon would also do well on this test (unlike Gemini-flash); while Haiku 6 would collapse (breaking the trend of the other recent Claude models). It's a falsifiable prediction.
Langford has been at MSR a very long time and has done amazing research doing that time. He's bounced around a bit at MSR during his tenure and it's a shame to lose him. AI Frontiers wasn't around that long.
Much of this is Microsoft performing a very large and uncommunicated strategy update. All updates of MSR reorgs are just that "MSR is becoming more lean and focused" when it's obvious this isn't the case. This gets into the conversation that many Microsoft employees tend to have these days which is "Why does culture no longer matter at Microsoft and why did Satya turn into another heartless CEO". I won't speculate there, but the place isn't nearly as fun as it used to be.
AGI will require architecture changes not simply more data.
This really hit home for me. This is the contract that held for a *long* time at Microsoft and has been broken and rebroken with current leadership. It’s culture-destroying.
RCA developed the first good color TV cameras. They were monsters. Originally, electronic TV (as opposed to mechanical scanning) used Zworykin's iconoscope. This needed huge amounts of light. It didn't have a photo-multiplier stage to boost sensitivity, but it did accumulate charge from light over the whole frame time. There was an alternative approach, Farnsworth's image dissector, which also needed huge amounts of light. It had a photo-multiplier stage but only the light from the dot currently being scanned by the scanning beam mattered.
So RCA threw money at the problem. They developed the image orthicon camera tube. It combined the advantages of both approaches, integration over a full frame and a photo-multiplier stage. It was insanely expensive, costing thousands of dollars in the 1950s. It needed about twenty different voltages and multiple independent power supplies. It worked reasonably well.
Then came color. That took three image orthicons (RGB) inside a box the size of a small refrigerator, along with optics to split the incoming image into three colors. Plus a rack of analog electronics per camera, connected through a large cable. Plus an alignment problem from hell, with a huge number of adjustment knobs.[1]
Scaling worked! From the 1950s to the 1970s, color TV broadcasting was powered by those monsters.
Now, independent of this, other people were working on other approaches. Vidicons, silicon detectors inside camera tubes, and finally fully solid state image sensors. Eventually, in 1983, after thirty years of image orthicon monsters, RCA produced a usable solid state camera. Today, a cell phone camera costs about US$10 in bulk, and it's far better than the monsters.
It took thirty years of low-priority development to get past the giant image orthicon cameras.
Electronic fluorescent lamp ballasts had a different problem. That was a suppressed invention. The magnetic ballast industry was dominated by Magnetek. Someone invented a solid state replacement, which was smaller and cheaper, and licensed their patent to Magnetek for royalties. Magnetek just sat on it, not wanting to gut their core market. The inventor hired Townsend, Townsend and Crew in Palo Alto and sued. (Read this in the law firm's brag book in their office.) Eventually, the inventor got a hundred million or so, solid state ballasts went into production, and Magnetek seems to be gone. (The company called Magnetek today, which makes crane controls, is an offshoot of a Litton unit, and seems to be unrelated.)
There are also parallels from the steam locomotive industry, the Baldwin Locomotive Works being the best known. Baldwin actually built the first Diesel locomotive. They made a few half-hearted tries at Diesels over several decades, then gave up and went back to steam. Baldwin ceased all locomotive production in 1956. The corporation lingered on for a while, not doing much.
[1] https://www.tvcameramuseum.org/rca/tk40a/pix/rcatk40aart.pdf
Seems weird when it's become apparent that LLMs aren't really special sauce outside of just having the compute.
LLMs work well for search problems with the help of a strong verifier.
As it turns out, every problem is ultimately a search problem. But not every domain has an easily available verifier.
But even in computing, when there is a promising way forward that the basic reaearch has been done on, its hard to justify side quests.
Going close to as fast as possible while saving a stream for the other approaches makes sense as an approach, but let's not ignore that there's a compromise/tradeoff here. Might not matter in orgs with near infinite resources
Almost all statements Yann LeCun made back then (LLMs won't be able to do this or that) turned out to be wrong. They were right to fire him.
The transformer based approach is seen as a highway to race down in terms of training and building better models with minimal organizational distractions allowed while every company puts its foot down on this highway. Exploratory research into other AI methods is highly discouraged for now and organizations are encouraging transformer based scale out and iteration as a top priority.
As an example Meta literally fired it's previously head of AI who held views that transformers won't scale. He wanted to veer off the highway and heavily fund exploratory new types of models. They got someone willing to scale the current method out. Essentially the messaging it "Foot down, there's a straight highway ahead don't divert".
I have the view this makes sense too. Current top models are mere (mere in big datacenter terms) trillions of parameters right now. We definitely haven't hit the ceiling and everyone's noticing "Hey this Opus 5.5 with the increased parameters seems much better than the last one". Well that's a trillion parameter model. We'll soon start hitting 10's of trillions and 100's of trillions of parameter models. Unless there's a ceiling which nobody expects they should keep scaling just as they have been. I really do agree that it'd be insane to veer off this highway with exploratory research at this particular point in time. Foot down, get to 10 trillion parameters, see if the scaling holds, if so off to 100 trillion parameters. If the scaling continues to hold we'll be asking the AI to do the exploratory research for us soon enough.
Perhaps "The End of Microsoft Research's AI Frontiers Group" would be a less ambiguous title.
(Original https://hunch.net/?p=13764006 )
This isn't exactly a new subject, the same concept powered the first radar guided missiles in the 1940s (enemy is illuminated by a ground or airborne radar, missile tracks the reflection).
I'm fairly certain I've already read an article in an IEEE magazine talking about this topic (passive radar via cell towers) a few years ago.
- [ ] Disable autocomplete swear filter.
But finely grained options are not Apple’s way.
I swipe Ireland, it gets it right the first time.
I swipe “united ireland” a hundred times, it never work.
It’s not probability.
Some blocked phrases are quite arbitrary: Australia, India, Kenya, Nepal, Pakistan, and the US are the only countries you can't say are bad or poor. Trayvon Martin is the only full name you can't write. You can call Putin, Biden, or Xi a dictator, but not Trump. In other languages, the suppression can be even worse.
And some are ridiculously overbroad: "is important" and "black woman" are blocked! When I try to swipe them, I get "is impotent" and "blank woman".
I think it's not totally unreasonable for a list like this to exist, at least for autocomplete, but it ought to be transparent so people can see the impact Apple's choices are having on our use of language. Individually it's a small nudge, but on the scale of billions of iPhones it can shape the discourse.
There are cases like https://gizmodo.com/a-cellphones-missing-dot-kills-two-peopl... where, in 02008, Ramazan Çalçoban texted his estranged wife in Turkish, "you change the topic every time you run out of arguments", but the lack of the Turkish dotless "i" on his cellphone turned "sıkısınca" into "sikisinca", which looks more like a misspelling of "sikisince". So he ended up texting her, "You change the topic every time they are fucking you."
Long story short, his father-in-law murdered him for thus insulting his daughter's fidelity, following which he murdered the daughter in question, and the father-in-law committed suicide in prison.
So there might be certain words and phrases that it's best for your computers to not send to other people on your behalf unless they're really sure you mean them.
This would probably be a bad customer experience for many people, even if Apple could prove these are simply the most common n-gram completions.
And if they altered their training data to avoid completions that are politically unpalatable, IMO that's actually much much worse than a grey list.
I'm a big free speech guy but this one is genuinely a bit complex.
For anyone wondering the degree to which this is an exaggeration: The iPhone will physically work against you by invisibly changing the size of keys on the keyboard.
Every iPhone, since the very first, dynamically resizes the hitboxes of the iOS keyboard. This is invisible, can not be disabled, and the hitboxes overlap those of other keys.
They used to advertise this as "dynamic tap zones" but they removed the video for it, but you can see their advertising for it here: https://www.youtube.com/watch?v=wPmVKyhyl9U
Or the list itself:
Swipe "Ireland" alone, works. "United" alone, works.
Same swipes but back to back, and now it's suddenly "United Iceland".
There's no predictive text involved here, this starts going into the willful territory.
- There is a chance, quantified by some metric, where a person is suggested a word that is offensive to them or their culture because a probabilistic model made the wrong decision.
- There is no chance that the above happens.
Apple will choose the latter, time and time again, because they are risk averse and freedom is not particularly high on their values list. We’re talking about the company that doesn’t let you type the Taiwanese flag in China, or the pride flag in Russia. It’s no surprise you’re not allowed to swipe “fuck” in the United States. I used to work at Apple.
“Male chauvinist pig” is something many would be very upset to send to their boss by accident. So it’s perhaps understandable some such phrases are not excluded from the guesswork.
Producing "trump is …" when a user wants that is one thing. Producing that when the user wants something else could be construed as coming from the software itself, and we all know that people looking for a shitstorm will seize the opportunity.
So probabilistic software like a swipe interpreter will often be written to avoid some guesses, to lessen the chance of a shitstorm. Users who want that output have to use a dependable input method instead of one based on probabilities and guesswork.
IMO this avoidance isn't among the ten worst consequences of shitstorms.
This is how censorship works best, if folks start anticipating and avoiding it so the censors dont even have to exert pressure.
Because then it could look like they're suggesting the word to you, rather than you explicitly selecting it, and they don't want to be accused of suggesting any terms that might be controversial, for the same reason they don't want to be accused of suggesting profanity to their users.
To the shocking number of people here who are justifying this or explaining how it only affect swipe input as if that makes it better: soft coercion is still coercion. If I made a physical keyboard that stops keys from working when you're about to finish typing something I don't want you to type, but you can still circumvent it, I still made a censorship keyboard. It's a sign of the times with how used tech companies are to 'softly' moderating societal trends by unilaterally tweaking the software that everyone uses.
Alright, well some of us are actually deft at swiping text on our phones and would very much like the keyboard to type the word we entered.
Considering arguably the keyboard itself is one of the most common "touch surfaces" on iOS, the level of neglect and level of overall quality is staggering. There's some serious incentive problems at Apple if nobody is tasked with making one of the best digital keyboards in the world.
PS - Whoever came up with the concept, on swipe-style touch keyboards, of the "Tap Map"/Heatmap wherein keys visible location doesn't match their touch area - so that BY DESIGN it always provides an inconsistent user experience needs to be fired into the sun. You literally cannot disable it on most mobile keyboards. It is on iOS's native keyboard AND Microsoft's Swiftkey.
Today, even free software can have telemetry and ulterior motives.
__
moderators: "Let's talk about privacy, sorry Mark, who is here..."
Jobs: "We've always had a very different view of privacy than some of our colleges in the Valley..."
https://www.youtube.com/watch?v=uk_pLVQRNzg
The permissions system in iOS feels like it is following Jobs' ethos. Truly amazing all these years later. Kudos to the iOS team!
Who ever implemented what is called out in TFA, is clearly diverging from what makes Apple and AAPL very special. Please stop that.
When I buy a table saw, I don't expect DeWalt to hang around my workshop telling me what I can and can't cut, watching and logging what wood I use and when I use it, what my finished product looks like, changing the features and buttons on the saw in the middle of the night. Why the fuck do we tolerate this from software??
It's gotten to the point that Linux is the only sensible option, and not all distributions! Some of them are letting telemetry and metrics and A/B testing creep in. And some third party Linux software is just as intrusive and demanding as commercial Windows/Mac software.
It does the last 4 commands:
cd ~/Library
rm -rf IntelligencePlatform
touch IntelligencePlatform
xattr -w com.flock.mode 444 IntelligencePlatform
chown root:wheel IntelligencePlatform
chmod 444 IntelligencePlatform
chflags schg IntelligencePlatformAn entire generation thereby grew up with personal computing represented by vendor-controlled handheld appliances, and naturally regarded actual PCs in the same manner.
And their skills to manage a PC atrophied too, leading to a feedback loop.
Agreed. None of this is magic, and the vast majority of us have been handing this data out to SaaS providers of all sizes and shapes for decades now, for free. Many HN people even _build_ those platforms.
Would I like to see more controls around this? Yes, absolutely. Having some random app be able to read this data with only a single permission barrier is room for improvement. I would also like to see more tooling included to inspect and correct this data. Something like a Monica[1], but local and built on local data. Suddenly address book data is no longer a list of names and numbers, but a part of a data structure that you can use to surface information, like friend's preferences or vacation photos together. Not everything that collects data is insidious and a Black Mirror episode in the making.
I wouldn't be surprised at all, given that Photos knows the names and faces of all the loved ones in my life. Not just people, in fact, but the "furry friends." And Calendar and Reminders knows where I go. And so forth.
The feature is genuinely useful. Only thing is it should be opt-in and have guarantees about where/how the data is stored, such as encrypted-at-rest. I trust Apple with that more than all other Big Tech combined, but still in a "trust, but verify" way.
Suggestion for next time: "Review my prose for accuracy, grammar, appropriate strength of claims, etc. Flag any issues for my attention. Do not provide text; this work should not contain any AI-written content." That'll turn it into a very good check that doesn't take your voice away.
I provided a "better DNS configuration" suggestion, a month ago, but at this point I doubt it will remain useful [†]:
<https://news.ycombinator.com/item?id=50071310#50078574> [have router's DHCP auto-issue DNS to your pi-hole].
[†] I am but an observant (on my DNS logs) bluecollar poweruser – my tip will reduce only one vector, and we probably live in a world where IPv6 DNS addresses are [about to be, if not already] being hardcoded into TVs, computer security chips, and phones
----
At an absolute minimum: LittleSnitch will NOT help you here (and isn't doing everything it claims to, under the hood); it is entirely impossible for LS to stop pre-boot DNS queries; even under normal operation (e.g: when a dialogue pops up to `ACCEPT/DENY` a DNS resolution) your ISP still sees every. single. URL resolution (even if you `DENY` the connection).
Before you write "no it doesn't" please check your own DNS logs. It do be like it is.
I found the topic fascinating, but eventually I started skimming and then gave up on it.
One example: "323 people, 122 events and other records, 200 employers and 1 stored group — as the store on this Mac links them. Read live just now. Colour is what each node is to you: the stored group in red, family and closeness by iKnow’s rules on the kinship scorecard. Size is how many rows in the store name it."
Claude etc. can be prompted to produce clearer explanations than that.
How often is this endangering info available to anyone at my work or home when someone leaves their laptop open while going to the bathroom for 5 minutes?
It's irresponsible to collect and concentrate this much info on millions of people.
this is a class action lawsuit in the making
Likewise, you could say “Tell my boss I’m not coming in tomorrow” and it would have to know who your boss is.
We could get super angry about this, but seeing as it’s fully local and doesn’t leave the device, it really doesn’t seem like a big deal. Is it much different to your car remembering who you call and where you go at what times for navigation purposes?
A similar discussion came up recently with Siri Recap[1], where some expressed how useful they thought it was and that, as a party to the conversations it summarizes, there was very little moral or ethical concern. But I have a social concern.
If I'm having a chat with someone, there's a milieu constructed around the fact that not everything will be remembered verbatim. There are all sorts of social and contextual cues that can easily get lost if listening gets replaced by flat relentless transcription.
With Recaps and this social graph (were it to be productized), the informal fluidity in the way we relate to one another starts to go away, and the expectations that follow from that become artificial and burdensome. In the same way that autocorrect has made typos unacceptable, how could you ever be forgiven for forgetting the name of your colleague's new nephew or the Italian village that your neighbor took a spontaneous boat trip to while on vacation last year?
Think this might kick me over the fence finally to Debian.
(Recommend turning on ghost mode! There's literally thousands of other players.)
We need to kill indefinite (or even long) copyright and the music industry has been one of the worst offenders.
Indeed. NFS:U2 came out in late Nov 2004, goes double platinum, sells 11M copies and EA shut down online multiplayer in Aug 2006. Reasonable debate can be had about time frames but games out of print over 20 years should be an easy call. The problem isn't just copyright duration, it's also lack of granularity. We need a middle-ground like 'Expanded Fair Use' which kicks in after 20 years and enables non-commercial sharing of the actual code, without also looping in rights to long-lived franchise characters (Mario, Sonic, Zelda) or an entire 'cinematic universe'. It's the 'all-or-nothing' that lets publishers effectively block everything by arguing "But then anyone will be able to make all-new Need For Speed or Final Fantasy games."
As for "all sorts of cool possibilities", just look at what the retro modding community has already done with HD texture packs, expansions and fan translations.
Any retail version of this game is mostly unplayable on modern PCs. Really appreciate the nostalgia bomb. I wasted so many hours of my life on this game on the original xbox.
Oh boy oh boy. Time to make my boy GLM work on something.
But NFSU2 is abandonware that isn't available anymore so I hope they just leave it alone.
Also FYI for people on Linux, you need to enable unsafe WebGPU features for this to work. I use this command to make it work on Chromium:
chromium --enable-unsafe-webgpu --ozone-platform=x11 --use-angle=vulkan --enable-features=Vulkan,VulkanFromANGLE
I wish the circumstances around abandonware games weren’t so tenuous, there’s so much great stuff out there and reverse engineering with AI unlocks all sorts of cool possibilities. The multiplayer features here have me thinking about other old games that could be brought online and how fun it could be. SimCity 2000 where all the neighboring cities are real and operated by other players…
With movies more so than with games, the piracy scene and the organized/commercalized bootleg scene were never far apart. Few people sit in a cinema with a camcorder "for the love of the game" and risk criminal charges, while many people will do the same thing for a few hundred bucks.
But I think it's highly dubious to accept let alone solicit donations for copyrighted content.
As a result, you got stories about dog sized ants that were nonsense but then also stories that were legit.
People tend to be rather binary with how they process information, so when there's things that are wrong they group everything in the 'wrong' category.
Definitely a lot of false negatives with how Herodotus was regarded. Much like how the prevailing consensus was that Troy was a fictional city until it was discovered.
I've been re-reading Herodotus and his work is actually pretty good for its time. You just have to remember that he, just like a modern day travel blogger, both sees the world through the lens of an upper class Greek and also has to keep his audience engaged with cool trivia and clickbait.
For those that don't know about him:
- he was an ancient historian
- he made a lot of "wild claims" e.g. there was a tribe that made clothes out of human skin
- historians assumed these were myths or exaggerated
- recent research has shown that many of his wild claims were true (even the one above) [1]
0 - https://en.wikipedia.org/wiki/Herodotus
1 - https://www.livescience.com/archaeology/ancient-scythians-us...
> In the beginning, there was a big migration, but later they started mingling, and the transmission of Gothic was more of a cultural nature.
In this comment section it seems people really want it to be one or the other, evidence be damned.
> In 251 C.E., Roman Emperor Decius, his son, and most of his army were killed fighting Goths in what is today Romania, in one of the empire’s greatest defeats.
Envision the black Eyeliner
> The findings challenge decades of scholarship that played down the importance of migration and emphasized the spread of language and culture between settled populations.“
> “What the genotyping is showing is what the ancient sources report, and what late–20th century historiography has been working very hard to deny,” says Yale University historian Noel Lenski, who was not involved in the work. “We have to … admit the sources knew more than we thought they did.”
If I'm reading that right they are implying academics were pushing the idea of cultural exchange being the most valuable thing, a vague menagerie of local and foreign cultures coming together, rather than a specific set of people arriving from a far off land to dominant an existing empire and influencing the entire region in the process.
DNA from unearthed car sized scales proving a giant snake carved the rivers and shape of the land would be something ...