HNHacker News
TopNewBestAskShowJobs

piokoch

3,832 karma · joined October 15, 2012

submissionscomments
piokoch··on Weave Robotics launches Isaac 1, a $7,999 home robot with Fall 2026 deliveries
Think of 90 y.o. lonely people who can't care less about some company seeing the interior of their house. Surely, it is a risk, but being without any help and assistance.
piokoch··on ArXiv's Next Chapter
That worries me a bit. ArXiv was and is great and so useful to humanity, giving access to otherwise closed knowledge, hold by publishers cartel, that I would not like to see it is turning into a "non-profit" of OpenAI kind...
piokoch··on TabFM: A zero-shot foundation model for tabular data
You don't need to. Just sample 1% of data, let the model do the feature engineering, ask model model to replicate everything in Pandas (or R, or whatever else), so you can run it for a full dataset.
piokoch··on Claude Code is steganographically marking requests
You are talking about "stealing" from people who already used stolen texts to train their models.
piokoch··on Professor denounces mass AI fraud on an exam at Brown
Thanks for the very interesting write up, I am teaching management studies programming Python (people who want to be quants, dana analysts, etc.) and I am struggling with the same problem - how the hell do grading to make it reliable.

LLMs has become good enough so whatever homework task I give, it can be solved easily by any LLM (and, in fact, this becomes unfair to those who are not able to pay for good model). I was asking to add comments & interpretation of results, this slightly helped, but LLMs are increasingly good in all this as well.

So, quite seriously, I am considering some on paper tests & quizes, because what else can be done?

Making people aware about importance of writing code is indeed a good hint to convince, at least part of the people to do this. Another thing is: they will leave university one day, they will search for employment and the employer might be much more hostile towards cheating during job interview and can easily make cheating with AI impossible... Question is, will it still matter?

piokoch··on Age verification is just a precursor to automated attribution of speech
Yup "protecting kids" has a long, long history of being used as an excuse to restrict citizens rights, yet something like Epstein Island was happening with all the proponents of child protection being happy visitors of the said island.
piokoch··on The best response to AI slop and online noise is from Robin Williams
One of the striking things about LLM-s is verbosity. A junior guy had to do some task and create docs how (a rather simple) change needs to be done. He produced, using Copilot (all those hyphens, additional lines between paragraphs, etc.) , a long dissertation that included some rather poor analogies and that could have been summarized in 5 sentences. Yet I had to go through that writing.

Same with the code. Generate Apache Kafka listener using Spring Kafka. Here you go, my human boss, the code is ready. The code is ready, but, somehow some outdated tutorial or Stackoverflow answer must have kicked in, hence it produced some totally unnecessary factory of factories that Java loves so much, but could be replaced with a few lines in the properties file.

But, when notified, Copilot kindly agreed that that factory is not really need and I am right.

piokoch··on U.S. government will decide who gets to use GPT-5.6
One more wake up call for anyone outside USA, especially Europe. AI will be weaponized, on the battle ground too, but the bigger battle will be fought in the industry competition. Those who have access to state of the art models will have advantage over those who does not.

Hopefully open-weight models will catch up, hopefully we, as the people, engineers will find the way to maintain those open-weight models on pair with the closed ones.

I try to be optimistic, as we won some battles, against all odds, Linux is flourishing, open source solutions are mainstream.

piokoch··on Too many R packages: CRAN is inundated with submissions
We have too many videos (since creating one is so easy), too many music (since recording it is so easy), too many books (since publishing an e-book is so easy). Now the same story happens again, for software. But this time it causes more troubles...
piokoch··on OpenAI DayBreak – GPT-5.5-Cyber
Soon, very soon, if you will need something useful, like medical advice, financial advice, you will be told that, well, ok, but you need to pay for an "extended license" that gonna be in thousands of dollars per month, otherwise you need to hire someone who paid that money.

The only hope are Chinese models, as Chinese commies are playing a different game as long as they are behind the flagship models (but it will change soon, like with cheap Chinese cars) and maybe, finally, Europe will start working on their solutions, instead of regulations.

piokoch··on Project Valhalla, Explained: How a Decade of Work Arrives in JDK 28
Yup. That's a big disappointment they could not cram universal generics faster. But I get the problem - they have to preserve backwards compatibility. I can take 30 y.o. Java 1.0 JAR and run it on Java 27 and it will work.
piokoch··on Feds freaked over Fable 5 after 'fix this code', not jailbreak, say researchers
There are big theories already born out of that glitch (like https://archive.ph/2OWwO#selection-1373.278-1377.12). The Doom is Coming!
piokoch··on Feds freaked over Fable 5 after 'fix this code', not jailbreak, say researchers
If it weren't for the IPO, Anthropic would just ship another model, called Opus 4.898, people would run another "duck on the bicycle" test that would be slightly better than the one from previous version 4.897 and move on.

But we have IPO coming, hence we face that big drama about model that would enable Iran to produce nukes, ok, that card was played, so maybe Taliban producing some magic poison to kill all Americans or some really bad people (Venezuelans?, Cubans? Somalian football referees?) to break into Github and make Github Actions working even worst (if this is even possible).

piokoch··on SpaceX to buy Cursor for $60B
Interesting, Grok, for a flagship AI contender was rather poorly performing. I mean, not bad, but visibly less capable.
piokoch··on Hetzner Price Adjustment
That's the difference between European and US companies and that's exemplification of the problem that Europe has. A big problem.

Firstly, Hetzner is really great, they have good service, good offers, they are rock solid, they shine especially in dedicated servers area - often better than all the cloud fad for many, many applications that does not need to scale crazily.

Having said that...

They expanded to a certain level and... just stopped. They do not have services that are making AWS/Azure attractive (all this identity/security stuff, MS Exchange like functionality), they are not even providing any viable messaging service, etc. Basic stuff.

As a result, companies who would even like to use them because they are solid, reliable, etc. simply can't, as Hetzner is missing basic services from business perspective.

So, they are not able to jump to the first league, have big customers, make big money, be able to invest into custom chips/infra, they are 100% dependent on US and Chinese providers. When something happens, like certain hardware shortage they are on the mercy of others and stop being able to compete.

Frankly, I don't fully get what the problem is. Luck of founders with vision, all those Jobs, Wozs, Zukerbergs, Elons, Bezosses? Luck of boring but effective CEOs (people like Eric Schmidt or Satya Nadella)?

piokoch··on Iroh 1.0
Still I am not sure why I should use their paid service instead of using publicly available infrastructure. If they go out of business, get sold what's then? DNS and friends are not going to disappear and send me "it was great journey" e-mail. Maybe for some specific applications, like P2P chats, this makes sens, but how many of such applications are needed?

I've looked at the usecases page, obviously there is an AI stunt (which I don't buy at all), for POS applications, well, there are better and less risky (see above) ways to do this, so the only thing that seems to make sense is this real-time sync, if someone is in the restricted environment (but, the point is, that in the restricted environment iroh is going to be blocked anyway by firewalls, z-scaler, etc.).

piokoch··on Ask HN: What are you working on? (June 2026)
Interesting, but given an easy access to AI, employers would get hundreds if not thousands of wonderfully written, properly suited CV. And everyone will have cool Github portfolio (with AI-generated projects). Good luck finding the right person in such environment.

So I am wondering what kind of tooling would be able to somehow spot the right people among flood of AI slops.

piokoch··on Claude Fable is relentlessly proactive
"When I came back a few minutes later I saw my machine open a browser window in my regular Firefox and then navigate to the dialog in question. I had not told Claude Code to use any browser automation".

Yup, tokens are eaten, money are paid. I am wondering how much energy/money is being burnt everyday by all of those LLM Agents on some useless activities like trying to recreate web application just to fix CSS bug.

And I would not call it proactive, proactive would be to ask for a CSS + HTML file in question, not trying to recreate them from screenshots.

piokoch··on Claude Fable 5
Obviously, soon, for anything valuable, you will have to buy from Anthropic "special license for biology/security/finance advises".

Question is if there will be any competition in this area...

piokoch··on Claude Fable 5
Monetization is coming. They'll tell companies, AI is replacing your workers, so it is still worth to pay 100K/year for the license, as those AI are not going to jump to other job, get sick, be late, complain, require free coffee and so on.

Soon the times of AI for $20/$200 a month will be long gone.

piokoch··on Claude Fable 5
"Without safeguards, Fable 5’s capabilities in areas like cybersecurity could be misused to cause serious damage"

What does it mean? That they have to add "safeguards" not do erase user disc, or, conversely, they are telling the audience that this model COULD be made so powerful to do some crazy stuff that can hurt governments, etc.? Are they showing off or threatening that if government X would not purchase the license the adversaries might do and what's then!

piokoch··on AI is slowing down
#3 Trillion, or even more, will be achievable. We are at the very beginning of the monetization of AI. It all started with the free Chat GPT, and a few others. Now the standard is $20 a month if you are not using AI tools too much and you don't need anything fancier. Otherwise you need to pay more, like $200 a month.

Unless you are not company and you don't have some "enterprise deal". And you need an enterprise deal, as 1) it guarantees that your (and your customers) data will not be sold to someone else 2) you are scared that your competitor will have such deal and become much more productive.

This is what we have now. What will be the future?

Well, soon, if you want something like financial advice or medical advice or job search/CV polishing you will be told, that your $20/$200 is not covering that, you need to purchase additional model to have that. Will you do that? It depends how much you are desperate to get medical advice or find a job.

Anthropic Mythos is an example. Soon, if you are programmer and you will ask AI Agent to spot a bugs, AI Agent will tell you that you need to buy extra model for this. Same with performance analysis, same with the design using tool X, Y or Z.

This is pretty scary, as it will put our well-being, productivity in the hands of few corps. It will be event worst that Google Search monopoly we used to have (until AI chats broke this, replacing Google Monopoly with a few other vendors monopoly).

Can this be prevented? Surely. Hopefully we will have capable open models and consumer-level hardware will catch up. But I think this is the place where governments should step in, invest into alternative models which will be at least comparable with flagships.

Chinese models shows that this is doable, DeepSeek is worst than Chat GPT/Claude/Gemini, but not that much and is clearly better than Grok (which is a huge disappointment for me). I guess India would join this game (especially with nationalist like Modi as the leader).

Europe could join this game, the problem is it kills its capabilities with high energy prices and inability to come out with some reasonable, well financed solution. So the only thing EU was able to come up with is some set of regulations that are blocking fast AI development in Europe...

There is French Mistral, but it is French, it is under-financed, it is only-French, as France would not like to lose control over it.

Germany have totally different strategy, they invest into manufacturing oriented AI, what makes a lot of sense, but does not help with the dangers we are facing.

The rest of the Europe is just too poor to spend billions on AI.

There is still time to buckle up for Europe, but given the course of events, stupidity of Brussels elites who does not see the storm coming I am not optimistic.

piokoch··on Siri AI
Why this is surprising? LLM-s are good in text generation on the base of the stuff they were trained on. Software is text generation, translation is text generation, LLMs can answer questions since billions were spent on tuning foundation models, that is people were collecting in (semi)automatic way questions with answers to the point we might think that LLM-s are "thinking".

Now people want to handle car rental. What are the relevant data that models were trained on for this kind of application? For Python code there is kirjillion examples on Github, for mathematical proofs there is endless stream of papers, books, etc. But for car rental? Mostly adds in the internet that want to trick you into a bad deal. So yes, LLM will be a disappointment, as it tries, well, to trick you into a bad deal. In addition, data are rather scarce so there will be a lot of hallucination, as it gets mixed up with yacht rental, bikes rental, ski equipment rental, etc.

piokoch··on I'm skeptical about efforts to revolutionize schooling
I was wondering about all this a lot.

I was teaching a lot of stuff to students: physics, math, statistics (during my university times) now I teach programming and Machine Learning.

I am torn between instructional based approach, which has this advantage that gives people a set of minimal skills to start doing stuff by themselves and the project-based approach, which is probably more interesting, but is very hard to squeeze in a relatively short classes time and also might left gaps, even in the base areas, as there is no time to cover everything end-to-end (think of teaching people about for loop, as it helps working with lists, but do not mention a while loop).

So, there should be some ideal holy grail in between both ways of teaching: show them everything versus let them explore and invent everything by themselves.

The crux is that instructional-based approach works great if it is well tuned to the student's needs. The problem is that every student has different needs and capabilities, so it is hard to do something that will work for everyone. So something is too difficult for some people, while being too easy for others.

That's why we have Bloom's 2 sigma problem - 1:1 learning works orders of magnitude better than in-class learning.

Now, LLM AI enters the scene, as the article is mentioning - individualized instruction could be finally achievable and I am much less skeptical about that than the author, as I tested that on myself, the good thing is I can ask and ask for more and more details if I am not able to grok something and my "teacher" is always patient, has as much time as I need.

It does not mean that teachers are not needed, just the opposite, because the key problem is to know what to learn, LLM will just do what you ask for, nothing more, so one need to know what to ask about. But once someone is on the specific topic and problem, you can really go quite far with LLM as a tutor.

piokoch··on Azure Linux 4.0 is Microsoft's first general-purpose Linux
I am not exactly waiting for Linux that will have obligatory ads and will take screenshots of my desktop and send them somewhere. Sorry Bill, but now, I've been through this already, I saw how superior DR DOS goes down because your mom was IBM board member, I had to use Windows 98 Millenium Edition, I was lucky to skip Windows Vista. So, again, no, thanks, never again.

Same with your cloud offering, ridiculous solutions like Azure Service Bus that has pathetic performance, pathetic API and high price.

piokoch··on United Airlines 767 returns to Newark after Bluetooth name sparks alert
... I can't believe what I am reading...

"Bluetooth speaker name had been set to a "four-letter word, [...] BOMB".

Luckily, it wasn't named "Nuclear Bomb from Cuba" because US Authorities would not have other choice than to nuke Cuba.

Seriously? What those people are doing when they see a fence with "ASS" painted on it? Do they believe that too?

piokoch··on The Last Technical Interview
Well, campfire method sounds easy in the blog post only.

You need an NDA from a candidate (many would not like to risk - what if someone gets job on the competing company and is sued for revealing some secrets from the campfire interview process).

In many cases interviewee would need to get a laptop from the company, as there are specific requirements about data security (disc encryption, usage of solutions like Fortinet or Zscaler), be added to company SAP to get access to the resources (Office, Teams, etc.), company need to purchase licenses for Office, Github, Jira, etc.

Surely hiring is hard, surely there are false positives and false negatives, but fixing this requires hell a lot of resources and organizational changes and costs.

piokoch··on Google employee charged with $1M Polymarket insider trading bet on search term
Yup, same with any kind of betting - sport, even draw games. There are obviously stories that someone managed to "game the system", like a man who figured out how to find winning scratch cards (Mohan Srivastava case) or Željko Ranogajec winning in Keno, but the point is that in the first case it was luck + skills, in the second it was overcoming the TOS by creating a lot of fake accounts, that's why the guy had to give his win back (details of the agreement were not revealed).

You bet against skilled people who set the stakes, so, yes, by observing numbers you can win in Keno, but if you comply to the TOS you will not win big money. The only chance to be able to "game the system" is to bet on something that lotteries brokers does not have time to look at, like 3rd Bulgarian bocce league matches.

The problem is that you need to somehow become an expert in 3rd Bulgarian bocce league and the money which are there are generally small.

I was investigating this (again) when AI showed up, as in theory it makes easier some analysis, but the big guys are also using AI.

piokoch··on The Green Side of the Lua
I have to say I am surprised that Java is better than Go in terms of energy/time.
piokoch··on Where does next-token prediction leave us?
If this were true, nobody would be buying iPhones, nobody would be buying Teslas, in fact, nobody would be buying premium class cars.
← PreviousPage 2 of 28Next →