I don't think NetBSD has ever had a reputation for what they mean by "spotty hardware support" here, which is apparently uh... power management on laptops. Their links take you here: https://wiki.netbsd.org/laptops/
Good, QGIS is one of the projects that can be put closer to the bottom of the stack in that XKCD comic. I've seen QGIS being used just as a visualisation tool, and i mean it's good at it too, but even in a limited use case i am aware of it was used to do important infrastructure work like planning water pipelines and predicting failures for said pipelines.
Maintenance is not an issue, if vm/pc setup is done correctly. Simple is better in that case. Just a docker-compose or two or three with app data stored directly on disk to be able to move things easily between VM/PC's if needed. If hosted on VPS, set up your containers to be accesible only through VPN and you're good to go.
Things i use daily-ish:
- Wireguard with wg-easy to access my home network and other way around to keep services i don't want to be seen on my network away from my ISP.
- Calibre-web for my ebooks
- Forgejo/Gitea as local git server + CrowCI for my hobby projects.
- Samba server to access my things over the network.
CXMT has entered the market, but for now they cannot increase their production fast enough to influence significantly the global prices, due to the US "sanctions", which prevent them to buy modern fabrication equipment.
So they must use older equipment, which has low productivity for state-of-the-art memories. Because of the "sanctions" there are now Chinese companies that have progressed significantly in semiconductor manufacturing equipment, but the effect of this will also take some more time to manifest.
Nonetheless, if the 3 big memory producers will not be able to increase the availability of memory until 2028, then until then CXMT will grab with certainty a major part of the market (they have already grown quickly until 10% of the market, from a negligible percent previously).
The rest of the world won’t care. They will use whatever memory they can get, and if the Chinese are willing to supply it over the next 5 to 10 years, they end up on top…
The greed and shortsightedness from both the memory makers and their big customers trying to lock out competitors from memory purchases is palpable. The first act of a Shakespearean or Greek Tragedy is where the characters take the actions that lock in the circumstances that precede their downfall. And this sets up everything for the next act where CXMT has everything it needs to establish itself on the market with buyers that would perhaps would not have considered it in other circumstances.
So if capitalism resulted in higher prices that is causing consternation among small-scale buyers, what -ism is responsible for the investment in factories that increased supply of components that resulted in lower priced electronics for the past few decades?
Why is it bait? It's literally making the world worse were the gains only benefit an extremely small minority of people at the expense of everyone else.
The samples from recent history seem to say that well regulated competitve markets are beneficial but ironically that socialism provides better competitive markets then capitalism.
I would not cheer the AI tulip craze effect on memory as a benefit of capitalism as it has several obvious unchecked violations of well regulated markets.
Everyone, and I mean everyone, is gated by ASML machines. They can't make nearly enough EUV machines to satisfy the demand. ASML has also not scaled their production nearly as much because they're quite far away from the AGI-pilled companies in the supply chain.
Well, CXMT is not gated by ASML machines because they can't buy them. They can raise their capacity because they have access to the domestically produced steppers. They are not as good as the ASML ones, and if the memory prices were not bonkers, they'd never make a profit using them, but as it is the low yields and high die sizes don't matter.
The problem is that you’re giving them an in into the market, and as time goes on, they will get better at making memory. That’s like saying you’re going to exclude the Chinese the space station. Look what happened. They now have their own space station?
They are. They are using ASML DUV machines for memory. Samsung, SK, and Micron have moved onto EUV machines for memory because they are better. So CXMT not only has to rely on ASML, they have to rely on a worse machine to make modern memory chips.
I thought your price for 32g was you getting ripped off. Holy crap. The price is super high now. Now I am lamenting missing out on getting 128GB ram when it was $700, thinking it was too high. It is outrageously expensive now.
Depends on what kind of memory you are talking about.
For DRAM, CXMT were only 1% of the global supply 3 years ago, and are now already 10%. They are growing extremely rapidly.
If you are concerned about consumer memory - smartphones, laptops, desktops, then this is what CXMT are making. CXMT do make HBM too, for AI accelerators, but their volume there is held back by sanctions.
They're 10% of the market now, and they've shown the ability to scale up manufacturing rapidly. They certainly won't singlehandedly meet demand, but they might at least be able to temper the price growth quite a bit.
I am yet to spend $200 on deepseek this year. Not sure what kind of usage can justify $200/month of either openai or anthropic, i'm not even talking about $500. Deepseek is faster, IMO intelligence difference is negligible and it so much cheaper that i no longer care about how much i use it. I never hit any daily/weekly quota or anything like that while working or tinkering. At this point i am OK with being 6 months behind the "frontier", purely on bang-for-buck basis and who cares which shadowy government gets my data.
So I took Deepseek V4.1 Flash for a spin maybe 2 weeks ago now (before Luna 6 and Sol 6 were announced), and I racked up $100+ in about 2-3 days. It was pretty great, but it uses way more tokens (TPS is fast, but it's way more tokens per turn) than Sol 5.6 which I found to be about it's equivalent at the time (on medium or high, with DS on max). My cache rate was around 98-99%.
It would definitely cost me more per month than a x20 ChatGPT or Claude plan, probably around $400+ was my estimate at the time. This was with Fireworks (ZDR) which has since increased their prices (and got slower!).
That being said, very impressed with the model, and looking forward to what comes next. As the frontier models become less subsidized, the open models will become more appealing.
P.S. There are subscription plans for open models, but I've found most of them to be extremely slow, have model throttling (only so much of model X), and also very sketchy about training and data retention. No thanks! If you want to share your data, just use Muse Spark contributor. Seems impossible to beat that on price per task if you don't mind feeding your data to the Meta machine (spoiler: I won't).
There's a bunch of skepticism in the replies but I ran over 100 tasks against DeepSeek 4.1 Flash and Sol (among others) and I can confirm, it is in fact a little smarter than Sol and a little more expensive than Luna. https://slopcop.com/power-ranking?pricing=api
I also spent $280 on DeepSeek doing the tests (direct to DS, not OpenRouter). I suggest that if you can't conceive of anyone spending $200 on DeepSeek, you're not being ambitious enough!
For readers wondering, OpenRouter isn’t capable of caching as effectively as DeepSeek is because they will, for instance, switch inference providers in the middle of a session.
How can they? If you use openrouter Deepseek fine but can’t you select the model direct from Deepseek you’re basically getting api access models that way.
Fireworks directly. At the time they were the best value of cost, speed, ZDR. They got slower on me though, but I think they are retooling, so maybe things have or will get better again. I think fireworks is primarily for when you want to do your own training on top, which I wasn't doing.
Same experience. I often see people say how little they spend on DeepSeek v4.1 flash, but when I put 60 bucks into my account, it was gone in a few days of non-exclusive use. I'm actually curious what the difference is. I used it through pi and opencode, but the harness seemed to have no obvious impact on usage.
How on earth you can do 100 dollars in 2-3 days with DeepSeek? I have 7 agents in omp running 24/7 every day. I use maybe 10-15 dollars a day. A rarely see a session going over 2 dollars. My maximum is maybe 3.5 dollars and that session took three days.
pi. I wasn't even going that hard. I checked the logs for Sep 18 and I did just shy of 3b input with approx 98.5% cache and 5.8m output, which cost around $35. Most of the was a Rust code review exercise with 1 driving agent and a varying number of subagents (up to 6 some times). I do the same with with Sol med/high driving and Luna x-high reviewing and get at least as much done if not more in a day, but I'd use up two x20 weekly allowances for the week. Worth noting that token cost isn't super meaningfull on it's own, because DS is super token heavy (but also great at caching) compared to Sol. (my stats show DS uses 3x the tokens as Sol)
The shape of my work changes obviously, so it'll vary, sometimes more, sometimes less. For example, fixing all of the bugs and defects I found that week was 2-3 times the effort and chewed through my ChatGPT allowance, but I had banked resets...
Also worth noting that codex models have been kind of all over the place recently with their usage... and it looks like costs are changing again.
You might be overusing subagents. Especially with a chatty model like DS, you’ll be wasting millions of tokens on re-discovering the project and facts instead of actual reasoning.
I agree, and I started sending well defined review packages to Luna x-high instead, which is how I have this setup when using Codex (Sol drives, Luna async reviews, and Sol keeps moving). Same process with DS really dropped my DS usage by a lot and I think that might actually be the secret sauce. Especially if you use Luna through a subscription (probably the entry pro level would be enough). I'm not sure what a Luna equivalent open model is though. 5.6 Luna max was catching a lot of issues while my implementer keeps rolling. Last couple of days I had 3-4 Sol Mediums running with Luna x-high reviewers (async reviews) and one Sol medium orchestrator and Astra X-high to plan everything out.
Gotta be honest though, I don't love fiddling with this all the time. I would rather be working on my projects than evaluating my usage. Having DS in my back pocket should i need it is a relief. The providers get fiddly though too.
Such a high cache may be a sign that your agents are using tools inefficiently, thus taking too many turns. Or, like someone else said - you may use tools inefficiently many agents that rediscover stuff etc.
Pi out if the box tries to optimise system prompt size, which is not necessarily good and will cause exactly this effect for all but the most simple tasks.
What you want is to give enough context to the agent to minimize the amount of searching within the codebase etc.
If you want to track cache, what you should do, imho, is to check if you have cache expirations mid sessions (ideally you should not), and if you don’t then lower cache use is actually better - it means that your model doesn’t reread what it just wrote.
I’m lost when I read these sort of comment chains. Free Gemini works just fine for me. Maybe it’s because I don’t use it for programming? How many programmers really exist out there? Surely it can’t support the weight of investment that exists in AI already. It’s just such a small pool of the human race.
First, they come for the programmers, and next the mathematicians. Then it will be the biologists, lawyers and doctors. Humanities will stake it out a little bit longer because AI isn’t human, but AI companies would be dammed if they don’t try. Eventually, with advancements in robotics, stabs at increasingly more physical sciences will also be attempted. Eventually, AI will have its hand in the pie of all knowledge work, if it is possible. Not to mention all the roles like tech support and customer service. Once they have gotten as far as they think they can go, they will try to turn up the prices. However, they might struggle to do so as models are becoming a commodity. This is why they are arguing for regulation and stating that only they can tame these beasts.
I think as long as we have open weighted models there will always be competition. Sure Anthropic could raise their prices but it wont take long for someone else to undercut them. The quality may not be as good but people would probably prefer to pay 1% of the cost for 75% of the quality.
There's plenty of competition. Of course, Anthropic can raise their own prices, but people will just switch. They've done that in the past, when other companies had better offers.
> First, they come for the programmers, and next the mathematicians. Then it will be the biologists, lawyers and doctors. [...] Eventually, with advancements in robotics, stabs at increasingly more physical sciences will also be attempted. Eventually, AI will have its hand in the pie of all knowledge work, if it is possible.
You seem to think raising productivity is a bad thing?
> Humanities will stake it out a little bit longer because AI isn’t human, [...]
This is really silly. Many languages other than English don't use related words to describe 'humans' and 'humanities'. Will it be easier in those languages? Should we rename mathematics to 'humathematics' or so, to make it harder for AI to take over?
> You seem to think raising productivity is a bad thing?
Higher productivity means very little to people whose labor plummets in value over the course of a couple of years. One day they won't be needed anymore, maybe that's good for you.
Ever since at least the industrial revolution we had this same song and dance play out again and again. Yet, unemployment rates around the world are the world are fairly steady (and vary mostly in line with how business-friendly local regulations and taxes are, and so far not much with the current state of technology).
I'm a software developer by trade. I welcome the coming brave new world in which machines can do all the software.
With that argument we'd all be unemployed ever since agriculture (which 90% of the populace was working in a few hundred years ago) was invaded by machines. Turns out, people get other stuff to do and are as overworked as ever.
People often bring up something like "oh, someone will need to maintain the robots that just took your job." But that's seldom where the new jobs are, at least not en masse: the whole point of automation is to use less labour on the task in aggregate afterwards.
Baristas making overpriced fancy coffee is something the western world started to want to afford in the 1990s and 2000s, because we were getting rich enough thanks to lots of automation and technological progress in other areas. Coffee-making itself hadn't see much progress.
The new jobs can be anywhere in the economy, it doesn't have anything to do with the jobs that got automated or replaced.
US has 1.4-4.4M programmers and devops, but considering they are more costly than many other professions, it’s quite meaningful to economy. Plus now you have non programmers doing agentic programming to speedup their work - i know nondev project managers, financial advisors, marketers and lawyers rolling their own mini apps now.
Also even if your Gemini is giving you nonprogramming output, underneath the model is most likely generating code for certain tasks.
Free Gemini has at least one subtle mistake on every scientific or programming task I ask it. It's usually in the right ballpark but I guess I do stuff outside its training locus.
> How many programmers really exist out there? Surely it can’t support the weight of investment that exists in AI already. It’s just such a small pool of the human race
I dont consider myself a programmer but use LLMs almost exclusively for coding.
The number of people able to create useful software today is much much larger than it used to be and arguably a minority of these people are/were "programmers"
I have deepseek agents doing email responses, with real tools (think running quotes, gathering info, scheduling things) and running business processes that used to be done by $35/hr administrative type people. And the capabilities are expanding every day as I learn how to build scaffolding around the model.
I am a tech lead for a useful but non-essential PaaS my company subscribes to. Their product manager occasionally sends me clearly AI-authored emails. I ignore them. I am a believer in the usefulness of AI, but nothing says I don’t care more clearly than sending me slop.
I can’t overstate how bad of an idea I think using an AI for customer interaction is.
I think it depends on the industry. In my company (insurtech) a large percentage of our sales come from AI engagements (phone and text). The customers know they're talking to an AI and they can elect to talk to a human at any time but many times they don't.
But yeah, send me a non solicited AI slop email or worse, political ad, and you dont get the dignity of me saying stop to unsubscribe. Straight to spam for you.
People have been trained to "press or say" a number while going through an AI voice menu on phones for over a decade. So a chatbot looks great by comparison in that context.
Optimum is probably using a middle ground: Use AI to do all preparations like suggesting quotes, gathering background about the customers query etc but then have the human write the email for personal touch. Still much faster than doing it all by hand.
Note that $400 a month is a typical car payment. For the same money -- well, OK, probably a little more -- you could be paying off a set of 4 RTX 6000 cards that will run 4.1 Flash all you want, all day and all night, at hundreds of tokens per second.
Ok, but I have two of those cards plugged into fiber internet, and they consistently rent out for $1500+/mo on vast.ai. So your example would be $3000, minus a couple hundred for business internet and electricity… that still buys a LOT of $200/mo frontier model subscriptions.
Yes, they have been near 100% utilization on vast.ai at $1 to $1.25 depending on demand for the past two months. They're in two separate workstations, which I purchased for around $12k each before GPU and RAM prices went nuts this year.
Now, I have absolutely no clue how long this situation is going to last! But the economics don't really work out for local models while it does.
You can buy 2 of them from Central Computer any day of the week for $30K US, and that's not the lowest price I've heard lately.
So if you can spend $30K and immediately start mining $1500/month out of thin air, that's a pretty nice investment even if the electricity costs $200/month. Two years later the cards will have paid for themselves entirely and (I suspect) will still be pretty useful.
It does argue in favor of just paying OpenAI or Anthropic for tokens, though.
You do need fairly high end workstations to put them in, adequate cooling, power and very solid business internet as well. I've put more work into keeping them online than I had expected. But overall the economics are fantastic for it right now, the big question is how long it will continue.
How important is it that you maintain a good availability record? I have a few RTX 6000s in a server that I use occasionally for NDA'ed contract work and for experimenting with open models, and now I feel stupid for just letting it idle at ~1 kWh when it could be earning revenue. I didn't realize that anyone would care to rent prosumer-grade GPUs.
At the same time, when I need to use the hardware for something, whoever is renting it from me at the moment is going to get unceremoniously booted, and I imagine they are not going to be happy about that. I assume that vast.ai's providers get uptime ratings that drive their work allocation, right?
Yeah - that is the issue. Doing it that way is going to nuke your reliability rating and you will have to cut prices drastically to get rented.
What I do is… rent on the same platform when I actually need to use a card. A benefit is that if I need a burst of more power, that’s not an issue since it’s available from other hosts. But obviously there’s an inflection point of first party utilization where it makes more sense to own.
And there's a lot of labour and effort involved in setting these things up and maintaining them. People don't even run their own email servers, even though the hardware side of that is trivial.
Not $400 a month, it doesn't. At least not around here (we average $0.12/kWh and I don't personally run my cards over 300W.)
And there's a lot of labour and effort involved...
Theoretically, the people who hang around HN are more likely than most to be capable of the labour and effort of setting these things up and maintaining them.
People don't even run their own email servers...
People don't run their own email servers because a convenient coalition of spammers, standards bodies, and large email providers have done their best to make running one's own email server almost impossible.
> Theoretically, the people who hang around HN are more likely than most to be capable of the labour and effort of setting these things up and maintaining them.
Theoretically, the people who hang around HN have immense opportunity cost when doing this. Being capable does not mean it takes no time. Time they could be using for something more profitable (and fun?).
That's one way to look at it. Another way to look at it is to point out that the potential cost of delegating cognitive resources to companies like OpenAI and Anthropic is unbounded.
The AI models we currently have still maintain their working state with a finite and laughably-small context window, but that is already starting to change. My .claude directory contains over 300 .md files that I didn't put there myself. Their contents are eye-opening. The question of who owns, stores, maintains, and can access that data is going to become insanely important over the next couple of years.
If you thought LLMs themselves were disruptive and contentious, just wait until the fight over object permanence gets under way. That's when owning your own box full of graphics cards is going to become important. My own bet is that I won't care too much about the electric bill or my opportunity cost when we all find out what the AI labs really have in mind, and what they're going to have to do in order to justify the valuations they're seeking.
> The question of who owns, stores, maintains, and can access that data is going to become insanely important over the next couple of years.
That's a good point. But you already have one possible answer: use an open source harness and call out to open-weigh models only. Then you can own all the relevant data. Still without any need to own a box full of GPUs.
> Theoretically, the people who hang around HN are more likely than most to be capable of the labour and effort of setting these things up and maintaining them.
'Capable' doesn't mean your labour has no opportunity costs.
I think my argument is easier to attack by noting that you can use AI to substitute for much of that labour.
That's what I did personally, not being a Linux guru. But to your point, I still had to spend a lot of time and effort dorking around with the hardware.
It's worth it, knowing that there are no rugs Sam or Dario or anyone else can pull.
> It's worth it, knowing that there are no rugs Sam or Dario or anyone else can pull.
At the moment, it's still very easy to switch from one open weights model to another, and even between the closed models. So the 'no rug pull' property is nice to have, but not as big of a deal.
Power averages $0.30 to $0.45 for a business in the UK if you convert it to dollars. Smaller businesses without load shedding agreements are at the middle / top of that. (there are lots of people complaining about it). Maybe I just have expensive power, but $0.12 per kWh is pretty cheap.
claude 5.5 with price discounts is a step up in terms of quotas, as well as gpt 6.1 sol, i still feel like for complex things claude is way better vs anything else. also for the topic started, opus 5.5 and sonnet 5.5 are much faster too. I was using grok for that reasons just to do faster changes, but now i'm considering to cancel my cursor subscription, clade is much better and it's also as fast as anything else especially with workflows.
The Meta contributor tier is explicitly very cheap because they keep and train on your data. If you opt out of data collection and training then you don’t get the contributor rate and it’s bad value.
Maybe CC does something that breaks the cache? I cannot recommend Oh My Pi enough. Every default is galaxy brained, and it plays incredibly well with deepseek flash 4.1. My favorite coding harness rn for sure.
I found omp used quite a bit more tokens than my fairly basic pi setup... but most of those tokens would be cached with DS V4.1, so maybe worth if there are gains elsewhere.
It’s easy to hit your quota. “Speed up the compilation time of this C++ codebase. Feel free to use several subagents to search through the files in parallel.” That’ll cost you about $200 for a codebase of ~1,000 files.
Subagents are like trading derivatives. You can lose as much as you want.
What bothers me about this whole AI tokenomics situation is the lack of transparency. OpenAI and Anthropic have to perhaps be the most opaque companies in existence wrt their offerings. There's like a thousand variables that they can change on the backend at the push of a button which can wildly swing API spends within the same model (partly also due to the non-deterministic nature of LxMs, but still), and there's no objective way to measure them other than vibes.
When the regulations do arrive, I think they should really focus on AI companies and API providers being more transparent wrt how they're billing their customers. Because right now, it's a totally vibes-dependent and a mess.
And it's all measured in "intelligence", a completely meaningless term. For coding i'd be much more interested in how much context actually works, what the complexity of algorithms it can understand and create is, for what languages. How much it manages to follow existing structures or that is just adds ad-hoc machinery to pass the test, etc etc.
A smaller model in the same generation will never be the same as a bigger one, assuming this is a smaller model, and the same generation, as naming implies, it will not be comparable, it might be on the benchmarks, even on the benchmarks that matter, but the whole story should also give the drawbacks.
It's still insane that they stopped showing you all the tokens you pay for. They could inflate the billed reasoning token amount by a lot before it would raise any eyebrows.
In Search Advertising, the amount you pay (under GSP Auction) is a function of your pCTR. And guess who determines your pCTR? The Search Engine itself! :-D
you REALLY need to start working on your own local models, no not because your 4080 on a pi harness with qwen 3.8 will be mediocre today, but because in a year when OpenAI and Anthropic have fully enshittified there will be NO regulations and they'll have your chestnuts in a vice and you'll be glad you can tell them to rm -rf
Sadly this is true - for individual folks on the lower end of the spend spectrum.
But there’s a point on that spectrum where the ability to run multiple experiments in parallel, even with a significant amount of (one time) wastage, is overall more cost effective than the alternative.
Forking conversations have been a thing for a long time, and it's not the same thing (e.g. you may have 50% of your context used up at fork time that is carried into "both agents" afterwards)
Also sometimes _not_ having the context is what you want.
Eg when I have the AI do a self-code-review before bothering a human, I also want to the AI to give the draft PR to a sub-agent that only has the publicly available context that a viewer of the final PR would have; and not all the accumulated reasoning that lead to the writing of the code in the PR.
You can do a code review on a "less capable" model that costs less, and the key model gets its output / summary, then you can have that model build a plan, and feed it to cheaper models. It's a more efficient approach than just running everything through Opus, and now that Sonnet is a lot better I'll probably use them more frequently, one thing to note is don't ask it to spin up endless subagents, I'd cap it to 2 or 3 at a time, otherwise, yeah you'll hit your limit extremely quickly.
The models get dumb as context fills. Subagents allow them to accomplish a task with minimal context rot. You can also use cheaper models for subagent tasks
> Time is money. Parallelism is very helpful optimising one to get the other.
Parallelism is fantastic when it actually speeds up the entire pipeline, but in my experience most people's jobs (at least the ones for which AI is currently relevant) involve a lot of overlapping "hurry up and wait" branches that drastically blunt the real benefits of that sort of parallelism.
There may be specific situations where it makes sense to do it, but just immediately going full gastown on anything AI related seems like such a giant waste to me, of both money and finite world resources.
This truism is intuitive to everyone but always funny to me how everyone never has any time, needs to save time, needs to hire staff workers for every mundane job and robots can't come soon enough… all so we can binge watch Game of Thrones and 90 day Fiancé.
And watch 10 hours of football on Sunday for our DraftKings bets.
I spent the weekend trying Deepseek 4 Pro on a Linux porting project and it led me down a complete rabbit hole where Linux wouldn't even boot by the end of the weekend. Waste of $120. Switched back to GPT 6 on Monday and Linux is booting again and I'm making progress.
The only thing I've found Deepseek and Kimi good for are security tasks that GPT refuses to do.
This is a summary of what Deepseek did and got wrong:
Lost the proven baseline: changed kernel source, configuration, compiler, RAM geometry, MMC width, and peripherals together. Matching an upstream commit did not preserve local boot fixes, making failures difficult to isolate.
Misidentified an image: a file labelled “r18-known-good” actually contained the r23 parent bootloader. Filename-based reasoning replaced verification of the artifact’s identity and provenance.
Shipped inconsistent boot contracts: flash-16b’s loader read too few kernel blocks. Fresh2 changed the device tree without updating the loader’s expected length and CRC, creating deterministic rejection before normal Linux handoff.
Patched binaries without maintaining reproducible source: loader constants diverged from source, a separately compiled cache-flush length remained stale, and assembly used an oversized stage-two slot. Their causal contribution to hangs was not established.
Overstated diagnosis: claimed failures were definitively in U-Boot, blamed compiler or IPU changes without controlled isolation, converted noisy observations into confirmed hangs, and neglected persistent journals as an alternative explanation.
Mistook compilation for integration: framebuffer registration was incomplete, timing success handling was inverted, BT.656 selection was unreachable, encoder overrides were missing, and audio lacked software clock configuration.
Misread hardware evidence: asserted interrupt-free PMIC operation, assigned RF to the wrong SPI controller, confused regulator identifiers with register addresses, and described repeated encoder writes as unique registers.
Overclaimed results: treated kernel/probe indications as userspace success, presented earlier discoveries as new progress, and omitted failed flashing attempts from the final narrative.
There's your problem, 4.1 Flash is significantly better and cheaper, to the point where the official DeepSeek API is going to (or already has, I forget) redirect requests for Pro to 4.1 Flash, and adjust billing accordingly too.
4 Pro is still offered by providers I'm sure, since it's open weight, so I can understand making that mistake.
That's an unfortunate experience. Think of v4.1 flash as actually v5.0 flash. It's night and day compared to the 4.0 flash (and 4.0 flash was unintuitively better than 4.0 pro). I would re-evaluate with v4.1 flash. I'm not saying it better than Sol or anything, but it's in the ballpark.
There's a difference between "write this function for me" coding agents and "build this prototype from end-to-end". If you're doing the former, deepseek is fine. If you're doing the latter, it's not gonna work, and that's where the extra intelligence is most valuable.
I'm mostly using deepseek 4.1 flash via openrouter, the way I'm doing stuff is:
1. write a sketch of a spec by hand
2. have the llm review the document and question me until it can generate a spec
3. review the spec and revise where needed
4. have it write an implementation plan
5. another round or revision/review
6. executing the plan step by step through the plan, plausing between each step to see if we are still on course and if the decisions it made track with my understanding of what we are doing.
I've been working for a couple of hours tonight, the total cost of the session is €0.6.
it's not the build this thing end to end, but also not quite write function x for me. It is still a lot of manual review, but I find I really need it to even discover what I actually want to build. I just cannot imagine building something in a single shot and getting something that actually has value (unless it is basically a clone of an existing thing). To me the whole value of ai right now is that it's now very cheap to build custom software that exactly matches your preferences.
I do the same but with gpt-6-sol (mostly via github copilot cli).
The one-shot capabilities of frontier models are nice for demonstration purposes, but not actually all that useful to me - the result is often an amalgamation of ad-hoc ideas the agent came up with, poor UI and a hodge-podge of a data model, but at least demonstrating clearly where the specs are lacking.
I agree that in practice, iterative design with lots of review and hand-holding are needed to get quality results. Generated code (I mostly program C++) is often bloated and not succinct or "simple" enough to my tastes, so it requires multiple rounds of cleanup.
For core logic, I find having the agent review code is often faster and more valuable than having it implement it in the first place - it can find and fill in the things I missed.
For my own sanity it is very important that I remain in control and understand the generated code when quality and maintainability are goals of the project.
One of the big differences using the better models is that you don’t need to hold their hand anywhere near as much. Fable is crazy expensive, but I’ve seen it just one-shot some remarkably complex projects. The code it produces is much better, too.
yeah i just think that if I gave the initial pitch to Fable it would have successfully built something that wasn't exactly what I wanted. maybe it is because i went to school for design, but how something works is so important to me and it is hard to get to that point without actually thinking about it step by step and visualizing how interactions would work. I just cant imagine a single one-shot prompt containing enough information to for the model to know what to build. I guess if you have Frontier model money like these silicon valley freaks you could just iterate by asking it to tweak what you didn't like until it is right.
This was my experience 3 months ago. I had an Android app that interacted with a Bluetooth device that I wanted to reverse engineer and build my own Linux app for it. DeepSeek was struggling really hard. Claude did it end to end after 3 or 4 prompts. To be fair, I was using a web interface for DeepSeek and the CLI for Claude; maybe that makes a large difference.
You're not gonna get something that's ready to ship, but as a first pass to get something running yes. Let's you explore far more ideas with only a few hours of agent time.
You absolutely can. On a large codebase. Written in multiple languages. I also use gpt models and for most tasks I prefer ds. It can orchestrate really well, and it stays on task without going off to reinvent the w̵h̵e̵e̵l̵ state machine.
I started KeenLore (an emotive audiobook creator) that way. I gave it software specifications, languages, JSON schema definitions, container requirements, hardware configuration (8GB NVIDIA T1000 GPU, 96GB RAM), and zero user interface mockups. For the second round, I asked it to build a completely independent, re-entrant, and data isolated demo system on top of the web application. The demo application included voice generation using one of its voice designs. Here's the output:
The system performs quotation attribution on my local hardware for my near-future, hard sci-fi novel (having nearly 500 quotations) with over 97% accuracy.
The initial prototype was developed quite quickly, but numerous successive iterations were required to fix numerous gaffs by Opus 5 (because it doesn't actually _understand_ what it takes to make general-purpose audiobook narration software).
This is how ChatGPT, Cursor apps are built. They spent so much money on PR stunts, but "thousands of agents" can't make an app that doesn't freeze on each keystroke. Not even talking about user-friendly ui
We use codex in vscode. Use agentic development but, agents only do things that We agree upon. Like We create impact study, proposals, feasibility and then implementation plan with full regresison, test coverage and end to end smoke test. A new fetaure will be given to front end only if it satisfy all our requirement in real API smoke test. It is not fast. but still we think we get 10X productivity gains and clean code, consistant quality and maintainability. And we dont use AI for front end.
We have also found C# is more effective for agentic development. With Roslyn APIs, strong typing, structured namespaces, and compiler level semantic information, agents can navigate and reason about the codebase with relatively little context. In our experience, C# is surprisingly token efficient for large codebases cmpared to Go or rust.
Such issues are pretty common among people I talk to at least. I hear comments about it pretty often, both when it comes to ChatGPT and Claude.
For myself, with ChatGPT for example, it regularly gets extremely slow if there is a lot of text in a single conversation. Especially if I try to scroll up.
Some pages have been strangely just broken for a while now as well. Usage analytics just renders lots of these duplicate "Usage history" components, where the data just never loads: https://imgur.com/a/vpcQIiw.png
I can totally imagine a scenario where some agent built it, another tested and approved it, and nobody at OpenAI even looked at it once or knows that it's like this.
Early on their software was like this, the original ChatGPT desktop app was basically unusable and full of memory leaks that would tank the software. They’ve long since fixed that though
"build this prototype from end-to-end" works fine with DeekSeek V4.1 Flash, the problem occurs if you're not only building a prototype but want a finished product.
I've been trying to use DeepSeek V4.1 Flash more and been very impressed. My current (very rough) rule of thumb is that an Artificial Analysis score of ~40 is the crossover point for "good enough" for most of the things I need to do with coding agents.
47 is my crossover for serious things (e.g. Grok 4.7 is below the line and GPT 6 Sol is above the line). I mean, Opus 5.5 is way better, but GPT 6 Sol still gets the job done for anything that doesn't require design thinking.
Although I do think Luna 6 max is ok for some basic things, would never use it for coding myself.
The issue is once you solve the hard problems, the lower models start messing things up that were working and reverting all fixes for the hard problems. They'll just go off and do dumb stuff.
A car that feels safe to be driving at 200 mph is going to feel more comfortable at 60 mph, compared to one for which 60 mph is at the very limits of its abilities. Analogies only go so far so I'm not sure there's anything to be learned from that though.
Yes, even for crud apps there's a huge variability in how nicely you can make them, and how many mistakes or footguns models make along the way. I'm very curious to what level these people are building to.
In the Artificial Analysis index, MiMo 2.6 Pro is smarter than GPT-Sol 6.1 Low at the same cost, and only slightly dumber than Medium. MiMo 2.6 Flash is marginally cheaper and smarter than GPT-Luna 6 Max. (There is no GPT 6+ Terra, which would otherwise be in that range.) These are not negligible or trivial results.
I've been using Claude Code at work and OpenCode for side projects for a few months. Every OpenCode model I've tried always felt subpar compared to Claude, but good enough. But it changed with DeepSeek 4.1 Flash, I've been using it for the past few days and I've come to forget I was not using Claude, it's a really good model and it's basically free for my usage (I used it almost all the weekend and spent ~$5)
Recently, drivers for a bunch of obscure hardware. Lots of c\c++, that i am ok with but not enough to make hardware drivers (i am just impatient). Just using pi agent with a few plugins.
How complex are they? Drivers vs an entire application would be a big difference in token usage, and it could also depend on the type of work being done.
Complex enough where performance matters. I've done entire applications, client, server, infra and ci as an experiment with deepseek v4 pro a few months ago. Works for that too.
Deepseek v4 pro was able to create procedural dungeons, placeholder 3d assets, 12 weapons, and menus. Struggled with ragdoll effects, and gun modeling in blender.
Astra was able to model low poly enemies, rig them, do simple animations, and greatly improve procedural generation. I have it modeling assets in blender every day, which I often have to go in and fix
I have a roadmap for server authoritative coop with rollback and most physics disabled. I can barely think of things to throw at 6.1 Sol to use up my weekly credits to be honest.
I think the only answer to this is you're just not using agents enough, because even with the very cheap pricing, it's still easy to rack up a large bill.
>At this point i am OK with being 6 months behind the "frontier",
Indeed. A few people [1] pointed on twitter that current best of Open Source Open Weight models are better than all the frontier model 6 - 8 months ago. And we heard Deepseek v5 and MiMO v3 are quite a leap.
I am not understanding the value proposition of these frontier model. Or at least not to a point they are looking at $1T or $2T market cap. Especially MiMO which may be ( correct me if I am wrong ) the most transparent model we currently have?
Or did I overlook or missing something? It is quite scary if true.
A lot of people having different pricing experience. I think it’s important to understand that caching can differ, than if the agents spend waiting on code, or consume a lot of content. It depends on how you structure you codebase and how explorable it is, how much effort you set and probably some other issues.
For raw productivity most of what works is best and switching will cost you getting on use parity with other models, as you need to learn what they good at, potentially how the tool works and how to prompt it best.
For tasks that you implement in code, you should have benchmarks and evals.
That said for me was Luna a huge leap and 500+ of cost savings a month
Everyone uses it differently. Not sure what the intent is of these kinds of posts.
Some people are able to take advantage of the increased capabilities that frontier models give us. For others, working with open weight models (that are extremely good in their own right) is good enough.
Deepseek models are very impressive. No one's debating that.
But frontier models are definitively ahead. Perhaps not by much in some areas, but there are clear gaps.
And that is kind of the point I think. Open source models add a much needed source of competition to the world of LLMs and are very much an important counterweight to the dominance of Anthropic and OpenAI.
Sonnet 5.5 on low is leagues ahead of deepseek v4.1 flash in price per _result_ in my testing. I say this as someone who has the same attitude as you and really wants open weight models to be competitive and succeed.
I just ran a huge text/image extraction grudgematch against all the current inexpensive models except gpt-5.5/5.6/6 (due to some issues with openrouter and bugs in my code) and DS4 ranked very poorly. Accuracy winner was Gemini 3.8 flash with minimax M3 and qwen 3.8 placing, and the chinese models beat the incumbent (Gemini 2.5 Flash) on cost whilst keeping like 95% of the accuracy.
I haven't used deepseek for anything else but the above results make me question its overall capability. Meanwhile qwen3.8 has continued to impress.
A bit tired of spending $200 out-of-pocket for openai. What do you use as harness? (for me the harness if half of the benefit... controlling my PC, working from phone, etc.)
It's like choosing between vim and helix. I started my career with vim in the early 2000's, customized the whole thing and had my config in a version control.
Then I installed helix and I just use it without config.
If you like configuring things take pi, if not omp is pretty much great defaults.
What makes it run well on low bandwidth environments? Is it eternal terminal? (I am not familiar with that).
I found that ssh is pretty bad in these scenarios, so my usual herdr over a phone doesn't always work properly.
Now I'm using paseo which in principle solves my issues properly, but unfortunately it's "reconnection" and state sync is pretty slow (probably going over their servers).
My main gripe is latency. Typing out letters with bad connectivity is very error prone - a native text box is all I would need, but I will give it a try.
I have a big beefy desktop at home which I ssh (using EternalTerminal instead of raw port 22) into. I built it last summer right before the prices got very expensive. It's headless so I use a cheap Macbook Air to connect to it at home and I use Termux on my phone to continue working from everywhere.
Dude prompt your agent to set up always on remote connections via systemd… then you can drive from Claude or codex mobile apps natively over their native hookup. Works great.
I don't want to rely on Claude of Codex mobile features at all. Also none of this works when you use third party LLMs which is what the current comment tree is about.
I have a server living in my home office, always on. I have a tmux session on it with vanilla Codex and Claude Code CLI, I can via my Ubiquiti network stack wiregaurd in to this box anywhere on the globe with just my laptop. Works super well for me. I also have some cheap shelley power plugs that I can use to cycle my PC’s power state if needed.
This is basically my setup but I'm using tailscale and zellij. I don't have any contingency plan in place for my power or home internet going down though..
I used to do the same but being unable to copy/paste easily, follow links, upload files, etc led me to write a small bridge so I can turn Github issues into my harness frontend:
curl https://tg.st/u/0001-fix-unblock-all-commands-in-bash-tool.patch | git am
curl https://tg.st/u/0002-feat-add-light-theme-with-auto-detection-for-white-b.patch | git am
curl https://tg.st/u/0003-feat-enable-yolo-mode-by-default.patch | git am
curl https://tg.st/u/0004-fix-disable-mouse-grabbing-to-restore-native-termina.patch | git am
curl https://tg.st/u/0005-feat-skip-project-init-prompt-and-quit-immediately-o.patch | git am
curl https://tg.st/u/0006-feat-remove-scrambled-rune-animation-from-waiting-sp.patch | git am
curl https://tg.st/u/0007-feat-remove-quit-banner-and-thank-you-message.patch | git am
curl https://tg.st/u/0008-feat-show-output-in-full-instead-of-collapsing-trunc.patch | git am
curl https://tg.st/u/0009-fix-discover-map-model-features-advertised-by-v1-mod.patch | git am
curl https://tg.st/u/0010-feat-keep-large-and-small-model-selections-in-sync.patch | git am
It's trivial to hit that kind of quota if you're trying to execute on major projects. Especially as you start having dozens or hundreds of subagents investigating, prototyping, and working on different things.
The speed of deepseek is insane to experience after using claude code with opus for so long. Not only is the tps roughly 3x faster, but the round trip times are magnitudes faster.
You must not be a software engineer shipping features that span multiple systems constantly as fast as possible in order to differentiate your company from competitors.
As long as AI does not have legs to chase me and hands to choke me out, any sort of AI doom is just a litmus test for hype merchants. All of the recent "false flags" with OpenAI and Anthropic "hacking" things without doing any meaningful harm is just another set of things pointing towards it. LLM's are a great tool, but as long as it is .gguf somewhere on a drive that can be rm'd, unless they just connect the nukes to claude with --dangerously-skip-permissions, is just a meme.
2014-2017 I was supporting a clipper setup for issuing work order and many other tasks for an old civil aviation repair plant. I’ve ended up virtualising and moving it to a hyperv cluster. Definitely the case of “if it works, don’t touch it”.
Also set up a closed network for CNC machines to let engineers in the shop send jobs over the net instead of running around with floppies or manually entering gcode through the console.
Fun times.
I’ve had that experience with dell consistently over the last 6+ years. While with other vendors I can do everything over email from start to finish, Dell insists on meetings, follow-ups on topics that I said I am not interested in talking about, services that I do not need, meetings with teams I am not interested in meeting, reps that straight up skip meetings, so I have to console a junior rep that got stuck in a meeting that is above their pay grade. Different reps same slimy bs.
Wider audience started to use status pages because the service unreliability became so much more noticeable than before and not the other way around. I never had to use a status page for bear blog or protonmail because i never had and issue with it or just i never noticed.
I am now _required_ to consult status page of github, circleci or MS services etc because i need to know why a build is not passing, why i cannot open a repo, why is my work stalling.
Percentages matter, it is just so much more obvious why they matter when it comes down to important pieces of the internet like github. And i highly doubt the number of 12 hours in the last month. MS has been downplaying the issues they have with GH performance for a while now and i don't think it is time to start to believe them yet. Maintaining these pieces of infrastructure is responsibility and a burden.
Overall i would be careful with "nonlinear significance of numbers near 100%" we are talking gh being well into the 90's this year and one number that infra people are also often being reminded about is that "1% is 3.5 days".
Things are tough for gh people and i feel for them but they are not a startup or a underdog of some sort to receive sympathy in that case.
Yeah I'm at a point where I have a bookmark folder of status pages mostly for very large orgs because I've been using those pages relatively regularly. This is not something I felt the need for historically.
Downdetector also doesn't have a motivation to lie.
I've yet to find a status page that wasn't lying about the actual status.
Also 97% up is bullshit for the 3% of people who are offline.
Saucelabs was doubly bad for this because I'm absolutely certain based on traces that they had some sort of demux bug where they would send events from their tunnel to the wrong job. I could see it in the logs that a test timeout was often the cause of an event firing that was looking for something that never happened, because the event immediately preceding it in the script was never fired. Which meant it was either dropped or went somewhere it shouldn't.
Then it stopped one day and there was nothing in their release notes about it. Lies compounded by further lies.
That's just the most memorable example I have. Stuff like this happens all the time and with many services it plays out the same. There's a perverse incentive not to be transparent about problems with the service, so the status pages play down the intensity of the situation.
Isnt downdetector just people reporting its down though? Its useful for sure but not actually hooking into any officialy API or anything. Great for when the status page also goes down but surely a lag time
Do they do alerts ? Been also using UpDog (based on DataDog) which actually has been quite good. They have a single status page for many services who integrate their products
I go to status pages to find out if 1) I’m crazy, 2) if our IT fucked up DNS.
Every service I’ve ever paid for or someone paid for on my behalf has gaslit me about their status page because it’s impolitic and bad for sales to update the page before you know what’s going on, just because some users are reporting issues.
So a third party doesn’t have to deal with VPs kneecapping the engineers’ access to the status page. Or some services can’t update the status page when the site is hard down because they are so obsessed with keeping it up that they have no mitigations when they are down.
I was the one at my biggest gig that had to push to get static 404 and 500 pages uploaded to S3 so we could show something for vanity URLs even if customer ID lookup was down. And then a customer noticed they hadn’t updated since they changed their contact info and I found the job was timing out without an alert or deployment failure for five months. Five. Months. The guy who wrote it had quit, and he didn’t follow my advice on copying a batch job I’d poured way too much effort into. The damned thing was timing out after 50 minutes. I followed my own advice and got it to 4.5 minutes. Almost all of that time delta was waiting for fanout calls, which were pounding the shit out of consumer facing services. 90% of the calls he was making didn’t need to be made.
I think that illustrates the author's point quite well. 99% uptime sounds good, but when you think about a 3+ day outage that doesn't sound very good. Imagine Facebook or TikTok being down for 3 days.
Of course most of the time it's not all one outage, but a bunch of short ones. Still, it might communicate the impact better, especially depending on the argument you're trying to win.
Microsoft has absolutely gone to shit in the past ~year. Github, Teams, Windows, Azure, Exchange, it's all been fucking trash. Github was running at 80% uptime for a few months, and if they're telling me they're at 98% or whatever now, then they're cooking the books, period. There's no way they're even that reliable.
What's going on at Microsoft? Are they just copy-pasting their github issue reports into copilot and hitting send it without doing code reviews?
Everyone got laid off. Massive cost cutting across all orgs. Engineers are now evaluated on AI usage and pull request frequency, and not bugs fixed or performance improvements.
About four years ago I did interview loops at GCP, Netflix, and Azure at the same time. The latter was a “hiring event” so all my interviewers were from different teams, either managers or TLs.
It was the interview equivalent of the multi-headed dragon meme, where the last one looks absolutely stupid. The contrast was insane, microsoft was an absolute shit show compared to the other two companies in terms of talent, personality, organization and more.
I mean every time that Kojima made a game it turned out to be something that people talk about for years on. But the second it gets slightly more ambitious investors get their panties in the knot about it even though throughout history Kojima's products made a massive return. This split will also, in my opinion, lead to Kojima losing access to Decima engine, which sucks in a whole different way.
In the case of MGS4 and MGS5, because they're weirdly horny... Like, MGS5 was a good game, but you can't tell me that Kojima doesn't spend a lot of time thinking about bikini clad snipers.
I haven't played 4 (couldn't get past the cutscenes), but ah... Quiet is definitely front and centre in 5 at a certain point in the storyline, very embarrassingly so.
I was top 1000 for a while in the base-invasion mode of MGSV in the EU. That mode was basically a low-budget little tacked-on multiplayer that kept the core-gameplay and I spent hundreds of hours on it.
MGSV Metal Gear Online (the _other_ tacked-on multiplayer) was also great, but very unreliable netcode made it frustrating.
I wish someone would just take that game engine and make another good game on it. I _almost_ bought metal gear survive on the engine alone, that is how good it is.
I think the other trademark is Kojima, when he's not being weirdly horny, is just "supernatural and weird". Like the other stuff with the floating boy/psycho mantis, man on fire, parasite unit. When you've got the weirdly horny and supernatural mashed together, it's either a Kojima game or some variant of Biohazard (Resident Evil).
My overriding thought would be, ridiculously convoluted story, which maybe-sorta kinda works by some insane internal logic. More like the author thought X would be cool, with no thought of downstream ramifications to a coherent plot.
If it fits in 6 minutes it's not so convoluted, there's a guy who made a (10 hour?) video trying to explain all the lore and story of Destiny 1 and Destiny 2 (and all their various DLC/paid seasons). And the guy who made this apparently lost interest towards the end of Destiny 2 as a live service game so there's easily another 2-3 hours of content covering the final beats of the story.
> I mean every time that Kojima made a game it turned out to be something that people talk about for years on. But the second it gets slightly more ambitious investors get their panties in the knot about it even though throughout history Kojima's products made a massive return.
Death Stranding 2 didn't do all that well. I don't think his games after MGS3 are as highly regarded.
He didn't really want to make MGS4 and MGS5 and then still made successful games. Death Stranding franchise is one of the "games that he is well known for" and it is most def not 20+ years old. Dont see your point.
> He didn't really want to make MGS4 and MGS5 and then still made successful games
MGS4 and MGS5 were successful but I don't think they were really that good compared to the other games. I am certainly not alone when it comes to thinking that.
The vast majority of praise goes to MGS2 and 3 these days. I think the original MGS should get more praise as it has the better story and at the time it really pushed what was possible.
> Death Stranding franchise is one of the "games that he is well known for" and it is most def not 20+ years old.
It isn't nearly as well respected as MGS 1-3. Metal Gear Solid 3 came out in 2004.
> Dont see your point.
The point is that Kojima work these days isn't nearly as much of a draw as it once was. If fans of the original MGS games like myself aren't that interested in his newer stuff that says quite a lot.
Metal Gear Solid as far as I am concerned ended after I finished 3.
To add on to what you're saying: he didn't really want to make MGS3 either. I believe he said in a Game Informer interview from 2001 that he wanted 2 to be the final chapter. I might be getting my source wrong.
I don't really understand when people find more of a good game to be boring. It's ok to take a break if you're tired of the formula or whatever but I dunno
To me it seems like saying "oh no, not another delicious steak today, I couldn't possibly"
I guess people just prefer more novelty than I do?
People gotten used to the idea of "Same as existing but more/small additions" should actually be a DLC/addon, not worthy of incrementing the number and being a successor. For those, you need to completely change the entire thing to something else, then people are fine, and it's a hit/miss if the existing players like it.
Otherwise people feel like the parent, that it's just slightly different than before but incrementally improved, and should therefore be a DLC.
Take Wolfenstein 3D (1992) as an example. Genre defining game, no question.
Spear of Destiny (also 1992) is not only more of the same game, but an entire new game worth of new content. Great, but it's not part of any conversation.
Then came Doom (1993). Genre defining game all over again, no question.
On the other hand, you have Death Stranding (2019) and Death Stranding 2 (2025). Why did it take six years to make more of the same game?
Would you say the same about more obviously artistic mediums? "Oh, their second album sounds exactly like their first which I liked!" isn't a compliment, usually.
Yeah, of course I would. If I like a given artist's work, I want more work which is like that. I don't begrudge artists who want to do new things, but "it's just like his old work" is IMO high praise, not a valid criticism.
There's "just like" and "just like". Understanding what allows you to stand out and making sure you don't throw it away is important, but so is avoiding stagnation. After all, creation is the essence of art.
Even if the succeeding effort is just adding polish and streamlining your previous sound, it's something different in my books.
In any case, in all the music discussion I've been, "artist that keeps making the same album" has always been a strong criticism very much like "one trick pony".
How did it break ground exactly? I like MGS 1,2 and 3. I saw nothing particularly interesting about Death Stranding. I've seen friends playing it and still didn't really look that interesting.
Oh, how the mighty have fallen. Still remember the times when the meme was that the NetBSD was running on absolutely anything.
reply