Reminds me of the SNL skit of TV newsreaders when the teleprompter goes down. By the end of it they've gone completely primitive and started killing each other.
SNL skits have these great ideas every now and then, only if the execution were a bit better, they'd have timeless pieces (like the Monty Python's dirty fork). This one was very funny anyway.
One show is produced with edits and retakes if someone flubs a line or breaks, the other is live. Also consider, SNL puts out more content in a single season with 22 episodes at 90 minutes, then all of Monty Python's full run of 45 episodes at 30 minutes. You're comparing apples to oranges.
Sometimes you look at an orange, and think how convenient it'd be if you could bite it like an apple. Okay, many don't, but few who do came up with all the modern fruits :)
You would think that 20 episodes a season over 50 years, then on just pure statistical probability they would have had a funny skit. An infinite number of monkeys on an infinite number of type writers will eventually write a funny joke. But somehow, against all odds, despite hiring comedians, comedic actors, improvers, and writers. They have never, once, gotten a funny skit. It defies belief actually.
Shouldn't matter right? You used Claude just to accelerate your production not to outsource your thinking. Its still all your ideas and stuff, so dont worry.
am i right?
Asinine argument. Are you outsourcing your thinking to the compiler because it gives you warnings, instead of getting an error two hours into your punch card programs's runtime?
People have done this work by hand for years and years, and now when they've been using LLMs to do it 5x faster for a couple of years, suddenly they'd be unable to work the traditional way? Please
> My car broke down and now I'm late for work
> Well you shouldn't have outsourced your mobility to a machine
One claude instance also doesn't need billions of dollars of operational infrastructure to work.
Its great that you can keep your computer running, 1/3 of my information is ithe cloud (which is not a problem at all, even today) and 100% of my infrastructure is in the cloud.
Its not a problem at all, that a service has an outage. My laptop also had an outage when the logic board broke. Took 2 days to fix so what
> My laptop also had an outage when the logic board broke.
> And if you had a local backup, you'd be able to resume your work immediately.
Immediately after buying a new laptop. Or perhaps you mean you keep a spare laptop around? In any case, it's also an object that requires a huge environmentally destructive supply chain to create. Quite like a LLM...
No. I could go buy with a older one, or in a pinch I could repurpose my current media server. Or I could put my data in an SSD with a bootable partition and borrow my wife's laptop...
> huge environmentally destructive supply chain to create.
Sunk cost fallacy? If the major electronics factories explode today, the existing computers won't stop working. If any of the Anthropic's datacenters go under, the compute capacity would have to be rebuilt.
If you're worried about the environment, you're better of stopping "outsourcing" your food gathering to factory farms or your walking to cars.
That's not what I'm talking about though, the question is whether the compiler does the thinking for you or not.
Or do you mean that the question of whether you're outsourcing your thinking to a LLM or not depends on whether you're running it locally? Would you claim you're outsourcing your thinking every time you run a script on a remote server?
> People have done this work by hand for years and years, and now when they've been using LLMs to do it 5x faster for a couple of years, suddenly they'd be unable to work the traditional way?
Yes. Not completely unable all at once, but they certainly get worse at it fast.
Losing skills due to over-reliance on technology is not a new phenomenon and not exclusive to LLM use, but they of course exacerbate it by so pervasively touching so many areas.
You take a clockmaker who's used to working with hand-operated lathes, mills etc. and let him work for a few years using CNC versions. For sure he'll probably be worse at using the traditional machines when he comes back to them, but that doesn't mean he's any dumber or worse at making clocks. He's just out of practice, same as if he'd have spent a few years staffing a convenience store or anything else.
That seems to be precisely what your study shows as well, though I only read the summary. Say the AI lowered the endoscopists' workload by (for illustration) 20%, so they got 20% less practice per unit of time, which made them have a performance that was 20% worse in the trial. I would expect the same to happen if they got assigned to do paperwork for 20% of their time as well.
I'm not trying to be argumentative, perhaps we're in agreement here, I'm just clarifying my point.
There's a famous talk about the importance of knowing when to use which level of automation, and maintaining currency at all levels you expect to need to use: https://youtube.com/watch?v=5ESJH1NLMLs
Analogies with physical tools don’t work well for LLMs because the former enable you to do things you couldn’t before, regardless how much you train and learn. A microscope and a needle allow us to see and work with precision beyond our human physical limitations. Furthermore they don’t make decisions, you are in control the whole time.
The talk looks interesting, thank you. I’ve added it to my list to watch later.
That only works because the clockmaker had the manual experience to begin with and only needs to remember to get back into the groove. If however he had always only worked with CNC machines the story would be very different.
I'm not convinced that's true. If you have experience driving an automatic car with lane assist, you'll have an easier time driving a 1980s manual car than somebody with no driving experience at all.
Likewise, the machinist still knows the principles of construction even if he's only used CNC machines. The endoscopist still knows the principles of colonoscopy even if he's only done it with AI assistance.
Just anecdotally, back when I took courses in assembler and logic circuits, there seemed to be no great difference in success between people like myself that had previous experience with that, and people who'd at most done some high-level scripting.
If you have enough curiosity I doubt using AI will be a roadblock to true understanding, just like high-level programming languages weren't.
Yes it matters regardless of how much thinking you offloaded. Having the tools you’re using vanish would stop anyone. Imagine if your favorite editor or other tools vaporized and you had 30 minutes left.
> Imagine if your favorite editor or other tools vaporized and you had 30 minutes left.
A bit off-topic, but "I" just created my own editor yesterday, maybe took 5 hours and it does everything I need, and I no longer need to swap between VS Code and Zed.
I think that ship has sailed. The entire world runs on SaaS tools which cannot / do not run locally. For mission critical ops, it can make sense to pay the enormous premium for local tools, but when one runs the risk/impact calculation, it rarely makes sense. Especially when there are four 9s attached to most of these online services.
> The entire world runs on SaaS tools which cannot / do not run locally.
Some of us have made different choices and therefore have different exposure to these kinds of failures. Yes, sometimes those choices mean higher costs or fewer features, but a lower probability of getting kneecapped by others' problems is a pretty good feature.
That ship is still very much available for boarding, but if your not already on it, it will take a bit of effort to get onboard.
Many of us refused to go along with the subscription model, and so are not affected. I've used primarily LibreOffice for ~two decades, and even more emphatically when Excel became primarily Subscription. Stuck with Gimp even after buying Photoshop for the Pantone support, but killed PS when it went online model. Was very interested in Fusion 360, but stuck with RhinoCAD when Fusion was online-only. I refuse to use online accounting/payroll "solutions".
I use Claude $20/mth a good amount, and am hesitant to use Chinese technology, but the last day makes it look a lot more sensible to setup local AIs, even at the cost of higher initial cost, lower capability, more system management effort, and potential compromise.
It's just one small shop, but many others could make similar decisions, instead of going with what's "just easier today".
If you use open router you can choose whatever you want whenever you want. Before I switched to mostly local I discovered like a lot of people how 99% of the time using the us labs is a complete waste of money.
I don't need to imagine it, this actually happened to me. OK, it was a bit more than 30 min, but I used Arch Linux back in the day, around '07. Coursework was due the next day and had loads to do. Was looking for a distraction so decided to `pacman -Syu` in the middle of the day. It completely borked my system. Didn't have enough time to fix so chucked in an Ubuntu CD and within an hour I was back with all my favourite tools.
It's very different when you own your tools rather than renting them.
> Having the tools you’re using vanish would stop anyone.
No. Not really. Depending on the presentation, it wouldn't stop professionals. More importantly, this is why you have backups. It's why you don't wait until the last minute to prepare your presentation, or, if you have to, you are knowledgeable enough to forego the presentation.
Simply put, in this case, absolutely not.
Edit: And to be clear, before you put too much weight on the "depending on the presentation" part, data heavy slides are not what I'm talking about.
No need to be so dramatic. There are a lot of things you can do with just speaking supported by a whiteboard with markers.
When I was still teaching, I’d do this sometimes. Just do a lecture with a whiteboard and draw data structures and algorithms. It’s often much more engaging than prepared slides.
Yes, I know it doesn’t work for every possible presentation, but for many it does.
Totally. But today, the room will not be well set up for it, markers may be dry, whiteboard covered in old, settled in diagrams, etc. All solvsble, but not in 30 min.
And for larger work presentations the expectation even 30 years ago was not whiteboards; that is very inefficient. It was transparencies. But try printing those in a pinch. And finding a projector.
That's not to say that a tool outage should kill a presentation (and if it was important, do it early); but it is not a do-not-care either.
Do you have a way to resolve the issue where the surface becomes degraded and much harder to erase over time? I think it's partly due to wear and tear and partly because people try to clean off the boards with solvents that degrade the surface that is erasable. Is there a way to renew it in some way?
The ones I have usually seen do have some coated steel (presumably because they are cheaper). At home we have a glass whiteboard (tempered glass I think?), which is extremely strong and the surface very smooth.
This is reminding me of a time in school were we thought if we stole all the chalk we might get a day off of Calculus class.
The professor came in and looked around for two minutes before shrugging and pulling a fresh stick of chalk out of his back pocket and began writing on the board.
This is the same professor who we'd never seen use a fresh piece of chalk ever, he always just soldiered through the nubs and remains littering the board's chalk tray.
We had no idea he actually kept a pack of sticks in his back pocket at all times.
Are people seriously so helpless they can't stand up and talk about what they've been getting paid to do anymore? If so y'all need ai detoxes, you arent even people anymore.
Standing up and talking is not the problem, it's getting the information into other people's heads effectively. The adage "a picture is worth a thousand words" does have a basis, especially when the subject is very technical.
Linus Torvalds might be able to describe some obscure part of Linux Kernel memory management that LLMs might not be able to come up with, but he won't be able to type it out faster than he can have it generated off his explanation
are your audiences going to be exactly as pleased with just the prompt in the center of the sole slide just as they would with the complete generated slides? probably not
will they find AI-generated slides sloppy and worthless anyway? I might, but that is not exactly my business. also I know gcc or clang won't accept a prompt as source file(s) yet
If you completely stop being able to work with without Claude, then I say a deep analysis of the root cause is needed and it may have some uncomfortable findings.
But for many, they use it as a productivity multiplier. They can still get the job done, just slower. The problem is that the rest of the supply chain likely doesn't care, management deadlines likely don't have 'do without AI' timelines any longer, so if we define the "job" as needing to meet the deadlines and not just complete the task eventually, people will find themselves unable to do their "job" without AI.
These both can be simplified as "not being able to do your job without AI", but I propose we have little to gain and a decent bit to lose by doing anything that treats these two situations as the same.
Few months I was in a similar situation, I ended up giving an improvised talk without any slides. I just talked, asked questions and it became a really interactive session. I also ended up showing websites and run into details people didn't understand. Afterward I heard people really enjoyed the talk!
I have a great feeling, it's cruisy as hell without emails! I'm able to get a morning of focus work done without interruption, and then I think it'll be a POETS Day.
I've not seen anyone use the term poet's day for many years. My dad used to be a tax accountant for British Rail and said that it was old corporate slang. (Piss off early, tomorrow's Saturday.)
Remember all those recent stories about "AI is really good at hacking now"?
Meanwhile SEGA's arcade machines have been offline for days: https://www.techpowerup.com/354714/sega-network-hack-keeps-j... and fanfiction writing site AO3 is down for two weeks so far. Then there's the freezer hacks, one of which impacted the French blood service.
It's been a long time since I read COMP.RISKS (and indeed it's no longer published) but this feels like an unusually high level of incidents.
It really is Friday the 13th today with outlook and now Claude both being mostly down. I wonder if this is related though. Either way not going to be a great ad for their IPO.
> Either way not going to be a great ad for their IPO.
Utility customer satisfaction goes up after power outages.
Obviously there are caveats to that. They must be rare, have decent communication, etc. But the point is that people will take always-on services for granted and be reminded of how much their lives are improved when they suddenly don't have them.
I got logged out of Claude in the morning, she couldn’t get back in without email. I guess Anthropic can keep their twenty bucks / days in November, for today.
Does every large company have a big server with some h100s and a couple of open weight models for just such an occasion. Especially if you don’t have the luxury of being able to wait or postpone eg under cyber attack?
I hope that doesn’t happen. There is a good alternative: if MOE models work for you, then Apple Silicon does a decent job at inference using whatever harnesses you like.
Good thing is at can expect some compensation such as resets. And yeah - diversify your providers, is the same with electricity from utility perspective…
My strategy: diversification. I run deepseek harness (which I made a vs code integration for and extended it a lot with plugins) and integrated it with my codex and claude subscriptions as well as an openrouter account. This way I can switch models and providers during my sessions and does not get blocked on such things.
Useful for
- outages
- running out of quota
- failing agents e.g. because of "cyber" flaging
Okay, this is slightly horrifying to me but also a bit genius...
But I guess Anthropic models are trained for Claude Code's tool call shape (i.e. the edit tool). Does it deal well enough with the MCP tools that Deepseek harness gives the model instead?
I don't think that "Anthropic models are trained for Claude Code's tool call shape". It's all about system prompt and model description (which is part of that prompt).
I observe that most models can deal well with the tools you provide.
Form time to time I analyze my sessions to see which tools could improve my (or my agent's) work. E.g. a background timer/watcher is on my todo list, so that I can chat with the agent session while its waiting for some shell response.
> Okay, this is slightly horrifying to me but also a bit genius...
If you're old enough to remember CGI web services, originally every request to a web server for non-static content launched a binary. Most of the web ran on that approach for a long time, later optimized with things like FCGI.
Good diversification. I like to have a strong local model like qwen3.8-flash-next available because I live in the mountains and my Internet is not 100% reliable. I don’t get it that companies allow themselves to get locked in on one vendor. Two or more, with local makes more sense.
dsh looks intriguing. Do they have any plugins that help effectively extend the context memory of each of the LLM's? I am starting to watch some videos on it but it will take me a while to get through them.
I think there are plenty ways to "extend the context memory". Compaction is the most obvious: reduce old context to it's core and free tokens. Another tool is session search. The models can look back in parts of the session that got "compacted away" and also in other sessions.
But honestly with 500k or 1m token sessions, I rarely run out of context memory and barely notice any loss of precision after compaction.
134 comments
[ 0.21 ms ] story [ 15.7 ms ] threadWish me luck!
SNL skits have these great ideas every now and then, only if the execution were a bit better, they'd have timeless pieces (like the Monty Python's dirty fork). This one was very funny anyway.
WhenI started watching, I was thinking I’ll share this on Teams to joke about the outage…then the third act came, and that it’s definitely nsfw.
People have done this work by hand for years and years, and now when they've been using LLMs to do it 5x faster for a couple of years, suddenly they'd be unable to work the traditional way? Please
> My car broke down and now I'm late for work
> Well you shouldn't have outsourced your mobility to a machine
And even in the case of outage, I can keep my computer running on a battery or a backup generator.
Its great that you can keep your computer running, 1/3 of my information is ithe cloud (which is not a problem at all, even today) and 100% of my infrastructure is in the cloud.
Its not a problem at all, that a service has an outage. My laptop also had an outage when the logic board broke. Took 2 days to fix so what
Of course it does. It's the marginal cost that is low.
> 100% of my infrastructure is in the cloud.
https://en.wikipedia.org/wiki/Turkey_illusion
> My laptop also had an outage when the logic board broke.
And if you had a local backup, you'd be able to resume your work immediately.
> And if you had a local backup, you'd be able to resume your work immediately.
Immediately after buying a new laptop. Or perhaps you mean you keep a spare laptop around? In any case, it's also an object that requires a huge environmentally destructive supply chain to create. Quite like a LLM...
No. I could go buy with a older one, or in a pinch I could repurpose my current media server. Or I could put my data in an SSD with a bootable partition and borrow my wife's laptop...
> huge environmentally destructive supply chain to create.
Sunk cost fallacy? If the major electronics factories explode today, the existing computers won't stop working. If any of the Anthropic's datacenters go under, the compute capacity would have to be rebuilt.
That's not what I'm talking about though, the question is whether the compiler does the thinking for you or not.
Or do you mean that the question of whether you're outsourcing your thinking to a LLM or not depends on whether you're running it locally? Would you claim you're outsourcing your thinking every time you run a script on a remote server?
Yes. Not completely unable all at once, but they certainly get worse at it fast.
https://www.thelancet.com/journals/langas/article/PIIS2468-1...
Losing skills due to over-reliance on technology is not a new phenomenon and not exclusive to LLM use, but they of course exacerbate it by so pervasively touching so many areas.
https://www.scientificamerican.com/article/how-gps-weakens-m...
That seems to be precisely what your study shows as well, though I only read the summary. Say the AI lowered the endoscopists' workload by (for illustration) 20%, so they got 20% less practice per unit of time, which made them have a performance that was 20% worse in the trial. I would expect the same to happen if they got assigned to do paperwork for 20% of their time as well.
I'm not trying to be argumentative, perhaps we're in agreement here, I'm just clarifying my point.
There's a famous talk about the importance of knowing when to use which level of automation, and maintaining currency at all levels you expect to need to use: https://youtube.com/watch?v=5ESJH1NLMLs
The talk looks interesting, thank you. I’ve added it to my list to watch later.
Likewise, the machinist still knows the principles of construction even if he's only used CNC machines. The endoscopist still knows the principles of colonoscopy even if he's only done it with AI assistance.
Just anecdotally, back when I took courses in assembler and logic circuits, there seemed to be no great difference in success between people like myself that had previous experience with that, and people who'd at most done some high-level scripting.
If you have enough curiosity I doubt using AI will be a roadblock to true understanding, just like high-level programming languages weren't.
That’s a gigantic “if”. Anything can become a positive in the right hands, but we have to consider the societal impact of broadly available tools.
If the OP had hours they’d probably be fine.
A bit off-topic, but "I" just created my own editor yesterday, maybe took 5 hours and it does everything I need, and I no longer need to swap between VS Code and Zed.
Some of us have made different choices and therefore have different exposure to these kinds of failures. Yes, sometimes those choices mean higher costs or fewer features, but a lower probability of getting kneecapped by others' problems is a pretty good feature.
Life is tradeoffs.
Many of us refused to go along with the subscription model, and so are not affected. I've used primarily LibreOffice for ~two decades, and even more emphatically when Excel became primarily Subscription. Stuck with Gimp even after buying Photoshop for the Pantone support, but killed PS when it went online model. Was very interested in Fusion 360, but stuck with RhinoCAD when Fusion was online-only. I refuse to use online accounting/payroll "solutions".
I use Claude $20/mth a good amount, and am hesitant to use Chinese technology, but the last day makes it look a lot more sensible to setup local AIs, even at the cost of higher initial cost, lower capability, more system management effort, and potential compromise.
It's just one small shop, but many others could make similar decisions, instead of going with what's "just easier today".
It's very different when you own your tools rather than renting them.
No. Not really. Depending on the presentation, it wouldn't stop professionals. More importantly, this is why you have backups. It's why you don't wait until the last minute to prepare your presentation, or, if you have to, you are knowledgeable enough to forego the presentation.
Simply put, in this case, absolutely not.
Edit: And to be clear, before you put too much weight on the "depending on the presentation" part, data heavy slides are not what I'm talking about.
Even better, if cave painting was good enough thousands of years ago it is good enough today.
When I was still teaching, I’d do this sometimes. Just do a lecture with a whiteboard and draw data structures and algorithms. It’s often much more engaging than prepared slides.
Yes, I know it doesn’t work for every possible presentation, but for many it does.
And for larger work presentations the expectation even 30 years ago was not whiteboards; that is very inefficient. It was transparencies. But try printing those in a pinch. And finding a projector.
That's not to say that a tool outage should kill a presentation (and if it was important, do it early); but it is not a do-not-care either.
(Yes, I know that the slippery slope is in jest.)
The professor came in and looked around for two minutes before shrugging and pulling a fresh stick of chalk out of his back pocket and began writing on the board.
This is the same professor who we'd never seen use a fresh piece of chalk ever, he always just soldiered through the nubs and remains littering the board's chalk tray.
We had no idea he actually kept a pack of sticks in his back pocket at all times.
are your audiences going to be exactly as pleased with just the prompt in the center of the sole slide just as they would with the complete generated slides? probably not
will they find AI-generated slides sloppy and worthless anyway? I might, but that is not exactly my business. also I know gcc or clang won't accept a prompt as source file(s) yet
But for many, they use it as a productivity multiplier. They can still get the job done, just slower. The problem is that the rest of the supply chain likely doesn't care, management deadlines likely don't have 'do without AI' timelines any longer, so if we define the "job" as needing to meet the deadlines and not just complete the task eventually, people will find themselves unable to do their "job" without AI.
These both can be simplified as "not being able to do your job without AI", but I propose we have little to gain and a decent bit to lose by doing anything that treats these two situations as the same.
We are all bosses now. Until it gets tired of working for us.
Few months I was in a similar situation, I ended up giving an improvised talk without any slides. I just talked, asked questions and it became a really interactive session. I also ended up showing websites and run into details people didn't understand. Afterward I heard people really enjoyed the talk!
So... be creative :)
https://news.ycombinator.com/newsguidelines.html
(I did log in and do 30 mins work later on once systems were up but we all have our moments of weakness...)
Meanwhile SEGA's arcade machines have been offline for days: https://www.techpowerup.com/354714/sega-network-hack-keeps-j... and fanfiction writing site AO3 is down for two weeks so far. Then there's the freezer hacks, one of which impacted the French blood service.
It's been a long time since I read COMP.RISKS (and indeed it's no longer published) but this feels like an unusually high level of incidents.
Utility customer satisfaction goes up after power outages.
Obviously there are caveats to that. They must be rare, have decent communication, etc. But the point is that people will take always-on services for granted and be reminded of how much their lives are improved when they suddenly don't have them.
Unless you’re on a negotiated 99.9% contract, you accepted this.
Luckily I can sometimes forget to check my email for an entire day, and it nobody even notices. At least _I_ am not that dependent on it.
Also some safe browsing thing is offline.. we just click "disable it and continue"... people are feeling very safe indeed.
Useful for - outages - running out of quota - failing agents e.g. because of "cyber" flaging
Since Anthropic doesn't seem to like subscriptions being used in third party harness, do they actually enforce that nowadays?
Edit: CLI in headless mode is an official feature
But I guess Anthropic models are trained for Claude Code's tool call shape (i.e. the edit tool). Does it deal well enough with the MCP tools that Deepseek harness gives the model instead?
I observe that most models can deal well with the tools you provide.
Form time to time I analyze my sessions to see which tools could improve my (or my agent's) work. E.g. a background timer/watcher is on my todo list, so that I can chat with the agent session while its waiting for some shell response.
If you're old enough to remember CGI web services, originally every request to a web server for non-static content launched a binary. Most of the web ran on that approach for a long time, later optimized with things like FCGI.
But honestly with 500k or 1m token sessions, I rarely run out of context memory and barely notice any loss of precision after compaction.