HNHacker News
TopNewBestAskShowJobs

mordae

2,350 karma · joined April 15, 2013

submissionscomments
mordae··on DeepSeek v4.1 Flash
Since it has low activated parameter count but huge total parameter count it needs more tokens to move the relevant information into the context.
mordae··on Switzerland's Federal Government Is Replacing Microsoft on 3k Computers
Sure. Even more so the handful of patent holders in the US.

Want to ignore those patents? Unless you have nukes you're not.

mordae··on How Europe is killing makers and micro-entrepreneurs
It is already €20+ to send something from e.g. Czechia to France.
mordae··on How Europe is killing makers and micro-entrepreneurs
It is not. Big European business always get what they want. Which tends to be maintaining status quo.
mordae··on Memory prices climb 500% in 12 months
Except staying on the edge costs exactly the same as not, when you take the resell value of the components on the second hand market into account. The problem is that some can afford to lock their capital in the hardware (via one-time investment) and some do not.

R9700 will lose less than 30% of value in the upcoming 3 years. Give or take. Yet people buy ancient cards for way more than 30% of R9700's cost.

mordae··on Qwen 3.8 27B is excellent, but it defaults to overthinking things
I think that in this case there is also the problem of trying to transfer MoE-style reasoning into a dense model. I mean, MoE needs reasoning to walk multiple experts, but dense model already has all the weights. So when you push it hard to reproduce the MoE traces, you are effectively asking a small mouse to role-play as an anthill. Not great.
mordae··on Qwen 3.8 27B is excellent, but it defaults to overthinking things
It is MoE. It needs to engage multiple experts when the problem is complex or unclear. So you naturally see more of those simply as a primitive it learns to use to page in more diverse set of weights. Remember that each token is just 6 experts out of 256. So it literally needs to tell its router that it needs a different set the next time.

And this memory control primitive leaks into the reasoning chain, because it has no other channel for it available and we do not know how to train any other channel.

On the flip side, it tends to converge quickly, roughly proportional to the actual difficulty / clarity of the task.

mordae··on Qwen 3.8 27B is excellent, but it defaults to overthinking things
It needs to argue with itself to extract most of the knowledge embedded in the weights into the context. Asking it to synthesize ideas directly in a single go is simply unreasonable. And MoE models need to walk multiple experts to extract all the knowledge on top of that. So you need to give them the reasoning trace to first spill all the associations into.
mordae··on Qwen 3.8 27B
Yeah, it should be basically free. No idea why it is not. I guess KV cache taking up RAM and possibly bad business sense or amortized engineering costs, I honestly do not know.
mordae··on Qwen 3.8 27B
DeepSeek V4 Flash is natively FP4 MoE with very compact KV cache. Say 8 GB/s. Qwen 27B is about 60 GB/s at full FP16 precision.
mordae··on DeepSeek V4 Pro 0813
0731 is definitely tuned for coding. I mean: https://gertlabs.com/rankings?mode=agentic_coding

But it is also a decent translator from English to Czech in my experience.

mordae··on Smaller, faster, safer: running Kimi and GLM at scale
Of course. On the other hand if you see 8 % gap for similar workloads, averaged across tens of sessions, with the same underlying model, it becomes a pretty clear signal. And I do exclude first request per provider per session from the statistic.
mordae··on Smaller, faster, safer: running Kimi and GLM at scale
Yeah, I've noticed them recently on OR and whitelisted. Then I backed-off pretty quickly after seeing the cache hit rates. It was also rather revealing to see how some provider hit rates differ when you are using them directly vs via OR.
mordae··on DeepSeek-V4-Flash Update
I was just using it when it landed. It started reasoning more extensively from nowhere and precision went up a lot. It also changed its prose style for the better. Looking forward to weights.
mordae··on Handbook.md shows that long policy documents do not reliably govern agents
To run at decent speed, all models try hard to use only most likely relevant part of the context and most likely relevant weights (MoE) to predict the next token. Doing the math in full is unfeasible.
mordae··on Handbook.md shows that long policy documents do not reliably govern agents
Why would anyone think that models optimized for efficient context management, giving much more weight to a short sliding window, would attend to distant, heavily diluted tokens?

Plus the model's capacity to take more context into account and actually integrate it to the output is simply limited by the number of activated parameters. If you give it a playbook, you are forcing to choose it between attending to the playbook and the task at hand.

If you want to force it to work step-by-step, you need to present the steps one-by-one. Ideally with rules for the current step at hand and maybe relevant input again, depending on overall task size.

Why did you think models love to re-read files before editing them? It increases recall quality and thus edit precision and thus benchmarks.

mordae··on It's getting harder to focus every day
Please also note that ADHD-like symptoms can present from sleep Apnea/UARS-induced sleep deprivation and the probability of having Apnea/UARS rises with age. Smoothly for men and a with a sudden jump after menopause for women.

So if you feel like you are getting more ADHD-like issues, do consider a sleep study. Early tells are easy: you often feel like shit in the morning.

mordae··on Sleep regularity is a stronger predictor of mortality risk than sleep duration (2023)
Sleep apnea symptoms are initially similar to ADHD. Apparently a ton of kids gets mis-diagnosed with ADHD while they are slowly suffocating because they have UARS. Stimulants mask it.
mordae··on Sleep regularity is a stronger predictor of mortality risk than sleep duration (2023)
Chance is you are not breathing in your sleep. You have Apnea or UARS and are slowly suffocating. Easy tells: you look like shit in the morning and it takes hours for you to normalize, you sometimes wake up all sweaty, you tend to open window ASAP in the morning "to get rid of the stuffy air", you are confused and/or have headache in the morning.

If you want to make sure, get a sleep study done.

Why do you think you are avoiding the bed as if you were afraid of dying in there. Exactly! You literally are.

mordae··on Google loses fight over record $4.7B EU antitrust fine
> alleged anti-competitive practices

I'd say that with court ruling these are no longer alleged. Right?

mordae··on DeepSeek open-sources inference optimizations with 60–85% faster generation [pdf]
Nope, people seek it out because government tells them to pay taxes _or else_.
mordae··on GLM-5.2 is the new leading open weights model on Artificial Analysis
They do not and it sucks for certain tasks.

It also means that if they actually trained with vision, they'd be on par with Anthropic models as vision seems to improve model performance across the board even for non-vision tasks.

mordae··on LLMs Will Replace 8-Track Duplication Engineers
DeepSeek V4 Flash on default (high) reasoning had zero formatting failures, just genuine losses. Total score:

| DeepSeek V4 Flash | 97/100 | 41/50 | 41/50 | 92W / 102T / 6L |

Benchmark costed me just under $1.

mordae··on Peopleless economy? Not technically impossible
> I think may be the government. The population will have to pay taxes for their maintenance. But it will be vastly less.

Governments are no longer allowed to print money. Do you think they will be allowed to build unlimited supply of robots? You are funny.

All the robot factories will belong to rich and we will have to _beg_ to have some measly allocation _while_ we feed the police that prevents us from just taking them.

And politicians will mostly just say this is _inevitable_ for some reason or other.

mordae··on Peopleless economy? Not technically impossible
> Money is an example. Really, what is money? A piece of paper? Some metal? Some number on a computer? Could you put your finger on what money really is, especially in the modern world?

Debt owed to central bank. Enforced by State via taxation and confiscation of property if you do not pay up. People seek it because they need to pay the said taxes and/or believe other people will seek it to pay theirs, including in a foreign territory. Loses value when central bank/State is unable to impose taxation and/or has not much useful work to extract from its subjects.

Recently western States have been captured by socialists, which prompted a reaction of the rich and powerful that eventually made the States unable to issue money and forced them to beg for the very thing they enable on the door fronts of the capitalists, making democracy a second-class economical citizen.

mordae··on The only scalable delete in Postgres is DROP TABLE
It drops dependents.
mordae··on The SpaceX IPO Will Be the Theft of the Century
Most have their pension automatically track index. Nasdaq changes rules so that the pension funds are forced to buy stock while very volatile and likely overpriced.
mordae··on Domain expertise has always been the real moat
> There’s no skill file that contains the tacit knowledge of a person who has reconciled a thousand payrolls.

That person has zero skill in actually making tight automation that doesn't just fall over. And I have yet to see an AI agent that tells them "look, your requirements are contradictory, given this and that, these two cannot coexist".

Those little sycophants will just go and try to please the domain expert and placate him in all ways possible. Bend backwards rather then forcing them to reassess their assumptions.

mordae··on Claude Opus 4.8
This is a terrible benchmark. It literally tests the models on their ability to track shifting line numbers. If they cannot keep up, no amount of abstract reasoning can redeem them.
mordae··on Can we have the day off?
Foraging was so good we only had to work like 2h a day. We had to start farming simply because the population growth from foraging was exponential and we have ran out of easy-to-forage ecosystems.

We've also ran out of freedom at that point. We used to simply walk away from bullies, but that was no longer possible, with the farm and all.

At some point bullies stopped charging for racket. They stole the land and told us we are now working under them, for them.

← PreviousPage 2 of 27Next →