At my very first job, we had a weekly meeting to assign tickets and estimates.
Our non-technical PM wrote "drag-and-drop" on a whiteboard, and then my name.
"Gavin, how long do you think this will take?"
"All week. At least. Maybe longer?"
PM laughed, then realized I wasn't making a joke.
"What do you mean 'all week'? That's ridiculous, do I need to give this to someone else?"
"Well, you could, but I wouldn't count on you getting it any sooner."
Appalled, he exclaims: "It's just drag-and-drop!"
After 10-15 minutes of a room full of engineers trying to explain the difficulty, he gave up with an unhappy "Fine."
The phrase "it's just drag-and-drop" became an inside joke for the rest of my time there...
Paper here: https://declarative-dragging.github.io/D2.pdf
IMO one of the remaining gaps of agentic coding is that it's hard for even multimodal LLMs to check interactive motion effects for correctness and "feel." Like, sure, they can take screenshots of a few snapshotted states, but in terms of how things react as underlying nodes move and (potentially) reflow. How do you QA the feel of moving your wrist rapidly in the context of nested zones?
The answer is that you avoid the need to: you have a (nearly-provably) correct foundation that ensures that behavior is predictable regardless of state, and just render accordingly.
The paper's approach is incredible here: simultaneous optimistic rendering of multiple possible states, and a composable algebra (a small combinator language) over how drop zones and distance are derived - allowing for an extremely high-level interface in practice. Everything feels snappy and right, and the code you write is about as minimal as you can get - though the paper notes that state management is somewhat complex. But this seems solvable!
Once this is ported to HTML, it'll become an incredible part of the modern dev toolkit. I'm very excited for what this means.
Being trans, of course voice affirmation training.
Then singing comes to mind, followed by language learning.
A solid visual demo like this would help give people more than "okay now more like the a in aunt"
- Quoted phrase in the opening sentence
- Opening quote: “if”
In the earlier days when they were slower and their code was worse so you had to review it, I’d watch the model code and just noodle away on guitar while it did its thing, interrupting it from time to time. Now they’re too fast for that to work, but still too slow for it not to be worth it to start a new task. It’s an awkward in-between stage.
If Anthropic made fast mode four or five times faster rather than just twice as fast, I’d probably shell out for it sometimes just to be able to single-task productively.
Every time I find myself idling in front of a screen, I feel the need to find something else to do, a distraction to fill those 54s before the agent sends a notification that needs my attention. This isn't healthy, and it's not how the human brain works.
Long, uninterrupted sessions of deep concentration are what helped humans produce their best work so far.
> "The page is only a viewer — git is the record."
> "Then comes the part that matters most."
> "That closing is the whole trick."
Not sure how much I trust an article on fighting AI brain fry in which the author outsourced writing a handful of paragraphs to AI...I hope the irony is not lost on anyone else, lol
- Ordinary term used more than once in the same comment
- “idle” appears 2 times
So I limit myself to at most 2 or 3 things at once, and if all of those are in long running sessions, well so be it, I know I'll be deep back in it soon.
But it's tempting to start "just one more session", especially something like getting it to better profile and improve performance. That feels super productive, but is often just to stop me feeling "idle".
- Quoted phrase in the opening sentence
- Opening quote: “Starship can't/doesn't/won't work”
fortunately one can just run debugger; to stop it early
Oh no, not again, I just cheated on my wife with a chinese spy, it just "happened to me"
- Quoted phrase in the opening sentence
- Opening quote: “intelligence operation”
It's worse than that - his mistress was a spy. This carries additional concerns since it opens the possibility of blackmail or extortion.
- Quoted phrase in the opening sentence
- Opening quote: “Nvidia Board to Probe Spying Case After CEO Admits Infidelity”
"Nvidia Board to Probe Spying Case After CEO Admits Infidelity"
https://www.bloomberg.com/news/articles/2026-11-08/nvidia-bo...
We should stop using "solve" when we mean "releasing products to target". Nobody will ever solve this problem with LLMs as they are currently architected.
- Quoted phrase in the opening sentence
- Opening quote: “Managing Director”
Also AIUI both claude[0] and chatgpt[1] offer HIPAA compliant BAA’s.
0: https://support.claude.com/en/articles/13296973-hipaa-ready-...
As someone who has been on the receiving end of many of these tools, I feel 'improved' is doing some heavy lifting here.
Completely overhauled to be operationally viable, perhaps.
Thrown into the dark recesses of the backlog, more likely.
For fast emergency destruction of documents, the NSA has developed a procedure https://media.defense.gov/2021/Jul/02/2002755769/-1/-1/0/EME... where paper and oxidizer get layered in a burn barrel. "The ratio which was adopted was 65 pounds of paper for 100 pounds of nitrate" gives you an idea how much extra oxygen is required for proper combustion.
When my phone is sitting on my countertop and I say "hey siri, add eggs to my shopping list" I expect it to add eggs to my shopping list, as it has done so for years, instead of saying "you'll need to unlock your iPhone first."
When I'm driving my car merging onto the highway and I see furniture on the ramp, and I say "hey siri, call the local non emergency police department number" so that I can report it, I expect it to say something like "I have the [city name] non-emergency phone number, want me to call now?" Instead of it saying "calling the New Jersey police department" which is halfway across the continent from me, and then giving me no way to hang up the phone hands free forcing me to grab my phone from the console and quickly try to hang it up while I'm driving at highway speeds.
Anthropic is way more focused. That helps in a way that builds upon another: their product tends to be better and the people's perception of it as well.
But "much less"? OpenAI needs a month or two of good choices, progress & announcements, and that flips. On the other hand, Deepseek could blow both out of the water unexpectedly, ending the argument about who has the superior tech.
There's also the Elon factor. Anthropic has a good relationship with Elon (to the point where he discounted the compute sold to them by quite a lot) while OpenAI/Sama is the devil to him.
- Quoted phrase in the opening sentence
- Opening quote: “from family offices and high net-worth individuals”
https://www.wheresyoured.at/running-dry/
They got a bit of pocket change at this insane valuation.
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.
- Quoted phrase in the opening sentence
- Opening quote: “The unbearable slowness of being: Why do we live at 10 bits/s?”
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”.
Claude Code doesn't even have a basic 'enqueue' prompt feature. Even though Copilot CLI, Codex and many other harnesses have had this for a long time now. And the GitHub feature request for this was closed, by you, referencing a completely unrelated feature.
With e.g. VM or container sandboxing, we can say things like "If there are no vulnerabilities in the hypervisor or hardware, then the agent can only access these files, these network services, etc". Auto mode doesn't even try to provide any guarantees, so it's impossible to reason about. It amounts to just using the models judgment about what's "malicious" or not, a definition which is subjective, changes model to model, and can be influenced by attacker input.
Empirical evaluations of how many vulnerabilities manual vs auto mode let through are not super meaningful because auto mode changes dramatically with model releases, the type of task being performed, whether the attacker is making targeted attacks designed to bypass auto mode, etc. A security feature cannot rely on these sorts of factors if it's to be taken seriously.
Further, percent of malicious tool calls accepted is a bad metric. It's like measuring the security of ssh in terms of what percent of unauthenticated logins succeed. The only acceptable value is zero. It's pass/fail, not a spectrum. So both auto and manual mode should fail equally here. Further, the set of malicious tool calls admitted in auto mode may not be a subset of those admitted in manual mode.
Whereas that should probably be default on instead, and this should apply to other known sensitive folders. A simpler system like .agentignore would be even better.
I like auto mode but "the agent supervises itself" is a bad compliance/safety answer on its own.
- Ordinary term used more than once in the same comment
- “approve” appears 2 times
This is called "approval fatigue".
If you think this can't possibly happen to you, you might be in for a surprise. This is an almost textbook case of classic conditioning. You'll find yourself hitting that button faster and faster even while fighting back the urge. So a human is no guaranteed backstop.
This is the scenario that cc auto mode guards against. It cuts down the number of permission prompts to something manageable, so that your fingers don't get trained into clicking 'approve' on 'rm -rf' in the very eyeblink that it appears.
So maybe you feel it doesn't work in theory; but weirdly enough, in practice this particular approach is actually one of the better ones; especially when used in combination with other common sense measures where practical.
This is an absolutely crazy statement without also bundling proper sandboxing. Nobody should be using auto mode outside of an environment they'd be happy with an 'everything is approved' mode too.
Many of the problems when it comes to the manual approval feel like they stem from the expectation of using auto mode in the first place. After all, what's the point of improving the manual approval and safety matching if auto mode is the end all be all, and what's the point of training the model to output commands in a manner that can be parsed by a human if auto mode is always meant to be used? It's very rare for me to not see "command could not be statically analyzed", because the model keeps outputting bash golf.
- Quoted phrase in the opening sentence
- Opening quote: “assisted manual review”
- Quoted phrase in the opening sentence
- Opening quote: “always allow,”
So pressing “Don’t Ask Again” by itself doesn’t expose passwords by itself. If you commit your local Claude settings and in those settings you have manually approved commands containing secrets directly in the command and then you publish the repo, that’s exposed.
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 ...).
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”.
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.
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.
Also, I've been playing with speech models for the past 2 years, and for me anything below 8-bit is a huge drop in quality; so quantisation for LLMs and for other bespoke modality models, is not 1:1.
[1] "Reliability Scaling Laws for Quantized Large Language Models", TUM, https://arxiv.org/pdf/2607.10855
- Quoted phrase in the opening sentence
- Opening quote: “xyz tokens/s on my light switch”
Reddit is flooded with these breathless headlines “xyz tokens/s on my light switch”. It’s always quantizations below 4 bits per weight. Mostly useless.
- Quoted phrase in the opening sentence
- Opening quote: “that's too bad”
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.
- Quoted phrase in the opening sentence
- Opening quote: “complete and comprehensive”
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.
- Quoted phrase in the opening sentence
- Opening quote: “sufficiently major”
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.
- Quoted phrase in the opening sentence
- Opening quote: “Cloju.rs”
- Quoted phrase in the opening sentence
- Opening quote: “TUI”
After testing several options at the moment I'm using Orca IDE as my daily driver - effectively as a layer on top anthropic / openai
- Supports both anthropic and openai with subscriptions and can you swithc easily between those two + other LLms
- Its a bit hidden / non intuitive to setup but it does support running everything on remove host -> I run (95%) of my jobs on remote hosts that I can both access from my laptop + their native app via a wirguard vpn
- You can switch between chatview + TUI on the same conversation - very useful if you move e.g. from desktop session to mobile view.
- For repeating LLM tasks Orca also suports a kind of scheduler - but I prefer currently to have the agent set things up via chronjobs on the remote machine (not super happy with this ...)
Key reason I dont want to go fully ChatGPT/Claude is also I dont want to build a super strong dependency on those mega players and remain the ability to switch to other interefernce providers / models.
As LLM usage is already a core workflow in our company - a sudden unavailability of one or a massive price increase - makes it absolutely necessary to build out alternatives.
We all know the downsides of SAAS Lock in too well from the 2010s. In that sense Fool me once, fool me twice..
- Quoted phrase in the opening sentence
- Opening quote: “LAMP stack”
- 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)
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.
3=(sin(arctan(sin(arctan(sin(arctan(2))))))^−2 − (2)^−2
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.
- Quoted phrase in the opening sentence
- Opening quote: “Kept you waiting, huh?”
> 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!
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.
- Quoted phrase in the opening sentence
- Opening quote: “horrific code”
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.
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"?
“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...
It seems like many times it's open source but "just to look".
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.
Nowadays you would just "npm install drag-and-drop" but back then you often had to hand-roll solutions.
If anyone can dig up an ancient library on GitHub for drag-and-drop back in those days, I'd imagine it would be several thousand lines long, not to mention the absolute nightmare that testing something like this with a proper framework and accompanying testing library.