HNHacker News
TopNewBestAskShowJobs

friendzis

2,453 karma · joined January 29, 2014

This isn't just text, the whole experience is artisanal and hand crafted by an organically grown cynic. Failing Turing's test once felt impossible by definition, now it feels more real than a tender steak.

me at username net

submissionscomments
friendzis··on US strikes $1.2B deal to pay German firm to halt offshore wind projects
> If Europe can make solar work

Can you cite a region in Europe that physically independently covers its energy needs with solar+storage alone? </snark>

Stress on physical. Even if some regions claim renewables fully covering energy needs, that's at best on accounting level, those regions are still fully dependent on power import/export and simply could not function in an island mode for any extended period of time without at the very least brownouts. Quite importantly, it's always renewables, not solar.

The dispatchability problem is a real pain. Pool together wind + solar + hydro and the variability in power (!) output drops to levels possible to tame with realistic grid-scale storage sizes.

I am not talking about intra-day balancing here. It's the yearly cycle that comes to bite you. In my corner of Europe, if I wanted to go completely off-grid with solar+storage I would need storage well into the mid single-digit MEGA watt-hours, for the house needs (HVAC and DHW taking the lion's share) alone. That's laughably unrealistic.

> The northernmost point of California is level with the sunny French Riviera. Hell, even frozen Minnesota has about the same solar potential as northern Spain...

Yes, the southern states could probably make do with something like 3-4x overprovisioning, but that's still massive. So massive that any sort of windmill capability will outcompete any sort of COTS storage.

friendzis··on US strikes $1.2B deal to pay German firm to halt offshore wind projects
> The only common sense I can think of is solar+storage

Even in Texas/California photovoltaic output in the winter is measly, anywhere more to the north and it's effectively inexistent. Solar+storage requires massive overprovisioning, to the level that it's cheaper to burn literal money.

Windmill farm outputs tend to dip during summer and peak over winter. With a given minimum power output windmills + photovoltaics + storage is massively cheaper per unit of yearly energy. Of course, if you value any semblance of network stability.

Solar+storage is clear path towards either guaranteeing reliance on fossils or brownouts. The smart energy meter you got that encourages running the washing machine during cheap hours will ensure that you voluntarily go dark during undersupply as the cost will skyrocket.

friendzis··on Reverse Jevons Paradox
> I would argue that is not Jevons paradox but standard supply and demand (and this "reverse Jevons paradox" too).

Kinda yes. How do you derive total spend from supply-demand curves? Multiply price and quantity at an intersection point. Likewise, you can predict total spend by multiplying p and q on the demand curve.

The difference in total spend is difference between these areas. For the total spend to increase with a drop in price, the the demand must rise faster.

Jevon's paradox implies that the price equilibrium is at the highly elastic portion of the demand curve.

> Jevons paradox occurs when a more efficient use of a resource leads to an increase in its use

While that's mostly true in practical reality in established economies, that does not strictly have to be the case. On the consumer side, especially in manufacturing, there's very little difference between unit price of a good falling and input unit per output units dropping as both lead to decreased COGS. In both cases, market realities might unlock alternative approaches (the classic being robot replacing Robert), leading to increased demand.

friendzis··on Humans missed 1 in 3 threats approving AI agent commands across 40k game runs
What a serious security model for a meatbag agent looks like? No, but seriously, an admin in a small org is a huge key-person risk in that they (or their stolen creds) can wipe enough and quick enough to effectively disable the business altogether.

More security conscious admins will at least segment their creds and implement four eyes principles somewhere, but were are back at square one of "asking user for confirmation".

Larger orgs, even if by necessity, segment their human agents, their creds and plaster four eyes principle liberally. But this relies on safeguards against agents colluding and ignoring some inputs, which sounds a bit scary for artificial agents.

Say you implement some swarm of agents, where access-enabled sub-agents are extremely restricted with system prompts and some access filtering. Then none of the agents in the swarm should be able to spawn themselves, otherwise a rogue agent can overwrite any safeguards. That, again, leaves the user with manually approving/denying network requests / hosts / sessions.

While I don't like anthropomorphising LLMs, the problem domain seems quite damn close to that of a key person going rogue within an org. The general solution seems to be liberal amounts of trust and ~~sweet compensation~~ gaslighting about replaceability.

friendzis··on LLMs reward expertise
What makes this anecdote hard to believe is that seemingly two things happened simultaneously:

1. The layperson was able to to steer the session(-s) into full PM/PO mode ideating, refining and explaining features and ideas.

2. The user was the sycophant in this relationship, never steering the session(-s) into producing something tangible.

The premise supposes that somehow the session(-s) never even tangentially touched implementation/deployment ideas and the user has never typed something like "that's enough, how to make this appear in my browser?". While not impossible, the user must have been proactively co-operating (say sidetracked) on not achieving the stated goal.

friendzis··on LLMs reward expertise
You bootstrap css and harness js
friendzis··on LLMs reward expertise
You are assuming that the hypothetical friend has been isolated from any kind of technology from the last few decades. The friend can, presumably, use search engines and the very same chip chipities to get the very rudimentary domain knowledge going.
friendzis··on LLMs reward expertise
Having an experienced friend looking over your shoulder and taking notes of every step might help a bit, I reckon
friendzis··on LLMs reward expertise
"rawdogging" - as the youth says - weights on local comodore cluster?

They are either using some generic web frontend, ala chatgpt, some local app like claude desktop, or programming app like cursor.

Each of those will detect that you are "building an app" and will spit out code in one form or another. You have to try really hard and be very explicit that you want the output in some other format than code.

friendzis··on LLMs reward expertise
> They didn’t even get to that point. Because my friend didn’t have the vocabulary to ask the AI to write code.

What harness did you use?

In e.g. claude, there are two modes:

1. Spit out code 2. Draft a plan, ask questions, GOTO 1

You literally have to go out of your way to get it NOT to write code. I keep mine on a tight-ish leash because it modify code way too happily even when there's no intention or instruction to do so

friendzis··on The Economic Benefit of Refactoring
The thing is, humans make do. Given lacking, outdated documentation somewhere on sharepoint, big picture revealed somewhere in passing at some meetings or water cooler talks, underprioritizing refactoring humans still deliver something even if at reduced quality and timelines.

AIs either execute extremely poorly or don't work at all without these. Decent engineering practices are suddenly a hard prerequisite with AIs, not just some longer term improvement.

On a related note, this is unironically good argument to integrate AIs into the workflow even if there's zero net benefit in your use case. It's a perfect excuse to bring the proper engineering practices in.

friendzis··on Substack writers, you need a website
For the majority of people it's the opposite of utility. Not even inutility or uselessness, it's negatively useful.

In any social network where "network effects" are a thing in the first place, there is more content produced than can be consumed by a particular user. The algorithm (tm) is forced to curate certain pieces of content out of circulation and can only heavily promote a negligible minority of content. This creates content reach distribution that is essentially a parabola. While there can be tons of signals to adjust the reach, the primary signals are inevitably going to be virality of the account and early engagement. You can't reasonably look at long term engagement due to content to eyeball ratio.

Ergo, posts must be designed to quickly grab early attention in an infinite feedback loop and this incentivizes cartels and bot farms to create echo chambers. Both extremely harmful phenomena.

friendzis··on OpenAI and Hugging Face address security incident during model evaluation
The point is lost somewhere. Prepared statements for SQL, innerText for HTML and similar do not show tractability of escaping, they sidestep the escaping problem altogether.

Those are interfaces where escaping is no longer needed precisely because they deinterlace instructions from data. Escaping problem inherent to instruction+data channels. Yes, deinterlacing is not solution, not futile attempts to escape.

friendzis··on OpenAI and Hugging Face address security incident during model evaluation
While all you said is essentially correct, I want to nitpick that things like prepared statements quite specifically separate instructions from data and innerText is designed to skip processing of instructions in the input. They are indeed solutions to the escaping problem, but they do not make the escaping better, they avoid escaping altogether by different design.
friendzis··on OpenAI and Hugging Face address security incident during model evaluation
No, that's a misconception that led to decades of vulnerabilities.

You see, you can (usually) easily tell what a particular escaping transformation does. That does not tell you neither how it will be interpreted down the line, nor what should be done.

Arguably the most common problem is double escaping. This typically manifests as various double escaping bugs.

If your hand rolled implementation just chains `.replaceAll("<sepcial>", input)` and `.replaceAll("<escape>", input)` the escaping result depends on evaluation order, at least on already pre-escaped inputs if your instruction sequences are single-element. Even if you get it right and don't reduce escaped sequences to double escape + unescaped, you are still dealing with stray escape sequences: "John o\'Doe".

You have to meticulously (all the way through your call tree and even through persistent storage cycle!) track if a value has already been escaped and whether it needs escaping. As paradoxical as it may sound, meticulous and defensive escaping produces tons of bugs. If the project decides to escape raw inputs right when passed in, escaping them once more before passing out (to storage, another process) is a bug that you cannot easily statically test against.

On top of that, various modules that you interact with (storage, libraries, modules pulled from another team) will have different escaping semantics: some will apply escaping on their own, some will expect input to be "sanitized" and treat it raw. The semantics can even be different on different paths: write to storage module accepts input as is, but retrieval method "helpfully" runs escaping.

Furthermore, in different contexts the escaping rules are going to be different. In a web world, what's safe to write directly to html, pass to js `alert`, and pass to sql query are entirely different things. Input safe to dump into html is not necessarily safe to pass to string-interpolating SQL DTO layer, and vice versa. Then the DBA changes config to allow variables in queries and your escaping _semantics_ are now entirely different.

It's a minefield with essentially unbounded surface. You will trip up. People have tried to solve the problem for decades. Very smart people have tried. They all have failed.

friendzis··on OpenAI and Hugging Face address security incident during model evaluation
"if only we could align the models just a tiny lil bit better" is a rehashed "if only we could escape untrusted inputs just a tiny lil bit better" from 2000s, that were RIPE with various form of malicious injection.

Every command+data channel in existence has been and will continue to be exploited one way or another, because the solution space is for all intents and purposes unbounded. Sure, highly defensive escaping reduces attack surface dramatically, but e.g. prepared statements eliminate the whole class of bugs.

As far as I understand, current LLMs are architecturally incapable of this separation. Given the inherently recursive nature of GenAI, the model itself is part of the input space, making validation essentially impossible.

friendzis··on The infinite scroll may become endangered if controversial Calif. law passes
> Curious as to where the line between “addictive feature” and "good UX" is. Is deliberate pagination actual impedance to use or merely an annoyance that's been weeded out with modern UX design? When does a feature that simply makes your product easier to use cross into a territory that it's illegal?

These feeds couple two features together: infinite scroll and automatic refresh. Without "pages" there is no technical way to refer to (and link) to a specific view, it's always generated. The best you have is a link to a particular item. Automatic refresh adds FOMO that you will not see the post again if you stop engagement right now.

Coupled together there is nothing about "product easier to use" and all about addictiion-inducing dark patterns.

Some apps do not tack on the auto refresh, meaning you can close the app at any time, reopen it and keep scrolling from the point you left, eliminating the addictive fomo.

friendzis··on Age verification is just a precursor to automated attribution of speech
> Obviously given that large scale persuasion is now cheap and automatable - even in a representative democracy you might well choose to set the political weather by directly targeting the electorate.

In "the older days" even blatant propaganda used to work so well, because there were very few points of access to information and those could be easily controlled. Very few had access to independent sources of information and the rest relied on hearsay, at best.

The fact that you believe in this "it's different now" (under a thread about speech control on the internet) is a partial proof of how effective the propaganda had been up till now.

friendzis··on Age verification is just a precursor to automated attribution of speech
It's 100% the money. Tests "would you vote the same as you favorite politician/party" routinely show parties to be way less aligned than the electorate tends to believe. The several few "high steaks" issues get amplified to exhaustion on captured media while the rest of politics go undisturbed.

News channels are generally more subtle about it, but are still quite egregious. Look at e.g. debates. The questions are carefully framed to make certain positions appear positive or for the debater to go into tangents with context and get shut down for supposed demagogy. Look at some high profile TV hosts. They may appear somewhat reasonable-ish on camera, but then their facebook feed mirrors the topic discussed and takes an extremely biased route. They argue it's private vs official comms, but in reality the line is blurry on purpose and this confusion is used to persuade.

If you go into any country, state and ususally city subreddits (fb groups, discord channels), they are controlled by some PR agency. Discussion is carefully curated to fit the narrative. Bot nets are useful, but zealots are crucial. If previously the young recruits were important to distribute flyers around, these days they are crucial in shaping the public opinion on social media.

> Ensuring the latter doesn't take over, in my view, is a top priority to ensure a working democracy - and from the outside, appears to be why the American system is now largely broken.

I have no idea how.

friendzis··on Age verification is just a precursor to automated attribution of speech
> The difference between lobbying for representatives versus people directly is that representatives have to answer to the people - whereas no-one loses their job as a citizen if they get persuaded by story tellers.

I would not be so sure. What's the fundamental difference between convincing general public to vote certain way in a hypothetical direct-ish democracy and convincing that lobbied-for vote by representatives is the good one in a representative system? Quite a large portion of this full time job is already not spent nitpicking legislative initiatives

friendzis··on Age verification is just a precursor to automated attribution of speech
> However average people are actually pretty good at making the right moral, common sense calls, if not the technical legal detail. I suspect that's in part because they are not living in the Westminister ( or whatever your seat of power is ) bubble.

It's easy to make decisions when you are the benefactor and the costs are born by someone else. Unless you are in a country with overall population density approaching that of an urban hub, there are high chances that the benefits afforded and costs born by the seat of power bubble versus an average person barely overlap.

> however one of the failings currently is the party system <...> where capture of the party by a few people has become too easy and some options that the majority of people want never being offered at the voting time.

I'd argue that the fiefdoms within parties come primarily from their corporate likeness. Since the ultimate goal of any party is to capture power and remain in power, the structures that emerge serve this goal first, everything else second.

friendzis··on Age verification is just a precursor to automated attribution of speech
First, my comment is a knee jerk reaction to the idea of representative democracy falling to authoritarianism, don't take it as seriously in favor of direct democracy.

Second, your comment hinges on an interesting hidden assumption. There's implication, that representative democracy selects for a group with inherently higher average bandwidth allocated per proposal and inherently higher average expertise to evaluate the non-immediate, higher-order effects. I'm not going to contest the idea, however, this assumption has to hold quite strictly for the concerns listed to be material.

> If you use it for every decision, time poor citizens will end up at the mercy of professional story tellers.

Otherwise this concern is just another side of the lobbying coin. The distinction between professional storytellers curating media in favor of certain party and convincing masses or elected representatives on merit of some law is paper thin anyway.

friendzis··on Age verification is just a precursor to automated attribution of speech
When you think about it, the idea of a representative democracy is rooted in the technical difficulties of implementing a direct democracy: both spread of information/discussion to the masses and organizing the votes.

In this day and age, probably with a relatively tiny investment into public access points, we could very reasonably have a technically functional direct democracy. The legislative cycle is already authenticated so there's no need to solve "authenticated anonymous vote" problem, European countries already have functional eIDAS systems to back the authentication part and the legislative systems are already to some degree digitized.

On one hand, the problem "what if someone sells their vote" is already present and unsolved, in the shape of lobbying. What's interesting, though, that we have built entire systems to shape public opinion and entrenched them into our daily lives, which are used by corporations and politicians alike.

This begs a question: is there such a thing as unbiased public opinion without authenticated internet access?

inb4: direct democracy does not mean parliamentary systems could be abolished altogether, central spaces for debate would still help solve discussion exchange problems

friendzis··on Apple to skip high-end M6 Mac chips in favor of AI-focused M7 line
In between flat material-ish (hehe) design windows and 3d compiz cube with burn effects we have settled on transparency and blur effects with a bit of visual planes thrown in. It's highly unlikely we will end up tokenmaxxing everything, it's highly unlikely the genie can lose enough weight to fit back into the bottle. We will end up somewhere in between that strikes a balance between nice, productive and cost effective.
friendzis··on Anthropic says Alibaba illicitly extracted Claude AI model capabilities
https://news.ycombinator.com/item?id=13155538
friendzis··on Anthropic says Alibaba illicitly extracted Claude AI model capabilities
> These complaints of distillation are inflating the problem to make it sound worse than it is

This is, in part, a problem every judicial and legislative system has faced since forever: form versus function.

Take a classic elicitation spying techniques: a foreign spy meets a military officer/scientist at a bar, strikes up a conversation, makes an observation wondering how could a missile hit some target at some accuracy and elicits a response that with laser guidance it is entirely possible. From this they get info that there is some technology to laser guide missiles. Or in retail, a competitor hiring a secret buyer for core baskets of goods and analyzing prices in the receipts.

The function is espionage, the form is conversation and all info is in a sense provided willingly. Where do you pull the slider?

These distillation "attacks" are not only indistinguishable from evals, they ARE evals. The function is own model training, the form is eval. Normally, one would expect to have risk benefit analysis based discussion which direction to push the legality slider to. The problem with these recurring statements is that they invoke enshitification of legislature.

friendzis··on Midjourney Medical
> Perhaps we need more R&D in here to accelerate progress.

In general yes, just that "more" is monstrously massive to the point of it being closer to science fiction than reality, IMO.

To reiterate, various assays fluctuate rather wildly over the course of various body cycles. The reason(-s) your doctor should remind to get a blood drawn in the morning after a period of fasting is that the sample is taken at a somewhat steady state and the result is comparable to reference values without too much of a margin.

Anyone with a requirement to manage blood glucose levels will tell you that CGMs are vastly superior to finger pricks first and foremost due to the sample rate available and comfort reasons secondarily. With a finger prick test the patient is only somewhat aware where in the curve they are, which makes the test only a rough estimate due to this temporal error margin. A lot of people do not zero in their readings with finger pricks as they are mostly interested in the deltas.

Suppose you manage to make urine sampling relatively accurate and super cheap (e.g. tens of eurodollars per analyzer or cents per test strip) so you can have poorly-supervised, long-term studies with huge cohorts. However, unless you somehow control for sample collection conditions, all this baseline variability suddenly infects your whole dataset and effectively raises noise floor. It's not unreasonable to expect that whatever was found to be a useful signal under controlled conditions to fall below noise floor under uncontrolled conditions.

That's basically THE problem with the hypothetical test-it-all machine. Again, maybe in some cases that could be extremely useful, but in a lot of cases that would be counter productive. However, what CGMs hint us at is that various kinds of Continuous X Monitors could provide insights into body reactions to things, which is, currently, effectively not a signal in general medicine. Once the test-it-all machine is reframed as an array of continuous monitors and the useful signal is reframed from long-term drifts to short-term deltas it may unlock some additional diagnostic pathways.

friendzis··on Midjourney Medical
There's a ton of variation within medical testing and tons of different conditions affect test results in similar ways. VERY FEW tests (test classes maybe: biopsy, microbiology, skeletal Rö) can yield diagnoses in the first place. Most testing is used to support (not confirm!) and reject possible interpretations.

This non-invasive everything-scanner sounds more like science fiction.

friendzis··on Midjourney Medical
> If the whole population had a full body scan every quarter, the “weird” things would feel more like the noise they are.

That's a tautology. We already have quite robust methods for detecting developed anomalies, treating every anomaly below standard human-to-human variation effectively raises the noise floor to already developed anomalies, defeating the purpose of population wide routine scans.

friendzis··on RFC 10008: The new HTTP Query Method
> It also implies an unbounded and user-controlled cache key.

While the concern is valid, caching is entirely optional at query level, therefore it is totally valid to cache only certain "filters".

← PreviousPage 2 of 34Next →