"Results suggest that the ease of collective action induces more egalitarian behavior by individuals in a position of power and makes those without power less willing to accept unfairness."
This is why capitalists dislike unions so much, becasue they know this. Together we are stronger.
This is a childrearing technique, too: say “please do X”, where X precludes Y, rather than saying “please don’t do Y!”, which just increases the salience, and therefore likelihood, of Y.
I've spent a lot of time trying to get LLM to generate things in a specific way, the biggest take away I have is, if you tell it "don't do xyz" it will always have in the back of its mind "do xyz" and any chance it gets it will take to "do xyz"
When working on art projects, my trick is to specifically give all feedback constructively, carefully avoiding framing things in terms of the inverse or parts to remove.
oh it's wildly different. About 15 years ago I worked on a porn recommendation system. The idea is that you'd follow a number of sites based on likes and recommendations and you'd get an aggregated feed with interstitial ads.
So I started with scraping and cross-reference, foaf, doing analysis. People's preferences are ... really complex.
Without getting too lewd, let's say there's about 30-80 categories with non-marginal demand depending on how you want to slice it and some of them can stack so you get a combinatoric.
In early user testing people wanted the niche and found the adventurous (of their particular kind) to be more compelling. And that was the unpredictable part. The majoritarian categories didn't have stickiness.
Nor did these niches have high correlation. Someone could be into say, specific topic A (let's say feet), and correlating that with topic B (let's say leather) was a dice roll. The probabilities were almost universally < 10% unless you went into majoritarian categories (eg. fit people in their 20s).
People want adventure on a reservation with a very well defined perimeter - one that is hard to map and different for every person.
So the value-add proposition went away since it's now just a collection of niche sites again.
Also, these days people have Reddit accounts reserved for porn where they do exactly this. So it was built after all.
“There are two novels that can change a bookish fourteen-year old’s life: The Lord of the Rings and Atlas Shrugged. One is a childish fantasy that often engenders a lifelong obsession with its unbelievable heroes, leading to an emotionally stunted, socially crippled adulthood, unable to deal with the real world. The other, of course, involves orcs."
A retort you often hear is that prior technologies, like writing or the printing press, may have stunted our calligraphy or rhetorical skills, but they did not stunt our capacity to think. If anything, they magnified it! Basically, the whole Steve Jobs' bicycle-for-the-mind idea.
My issue with applying this reasoning to AI is that prior technologies addressed bottlenecks in distribution, whereas this more directly attacks the creative process itself. Stratechery has a great post on this, where he argues that AI is attempting to remove the "substantiation" bottleneck in idea generation.
Doing this for creative tasks is fine ONLY IF it does not inhibit your own creative development. Humans only have so much self-control/self-awareness
CS Peirce has a famous essay "The Fixation of Belief" where he describes various processes by which we form beliefs and what it takes to surprise/upset/unsettle them.
This blog post gestures at that idea while being an example of what Peirce calls the "a priori method". A certain framework is first settled upon for (largely) aesthetic reasons and then experience is analyzed in light of that framework. This yields comfortable conclusions (for those who buy the framework, anyhow).
For Peirce, all inquiry begins with surprise, sometimes because we've gone looking for it but usually not. About the a priori method, he says:
“[The a priori] method is far more intellectual and respectable from the point of view of reason than either of the others which we have noticed. But its failure has been the most manifest. It makes of inquiry something similar to the development of taste; but taste, unfortunately, is always more or less a matter of fashion, and accordingly metaphysicians have never come to any fixed agreement, but the pendulum has swung backward and forward between a more material and a more spiritual philosophy, from the earliest times to the latest. And so from this, which has been called the a priori method, we are driven, in Lord Bacon's phrase, to a true induction.”
Fundamentally, we are at a point in time where models are already very capable, but not very reliable.
This is very interesting finding about how to improve capability.
I don't see reliability expressly addressed here, but my assumption is that these alloys will be less rather than more reliable - stronger, but more brittle, to extend the alloy metaphor.
Unfortunately for many if not most B2B use cases this reliability is the primary constraint! Would love to see similar ideas in the reliability space.
> Ordinary language is totally unsuited for expressing what physics really asserts, since the words of everyday life are not sufficiently abstract. Only mathematics and mathematical logic can say as little as the physicist means to say.
- Bertrand Russell, The Scientific Outlook (1931)
There is a reason we don't use natural language for mathematics anymore: It's overly verbose and extremely imprecise.
I recently started diving into LLMs a few weeks ago, and one thing that immediately caught me off guard was how little standardization there is across all the various pieces you would use to build a chat stack.
Want to swap out your client for a different one? Good luck - it probably expects a completely different schema. Trying a new model? Hope you're ready to deal with a different chat template. It felt like every layer had its own way of doing things, which made understanding the flow pretty frustrating for a noobie.
So I sketched out a diagram that maps out what (rough) schema is being used at each step of the process - from the initial request all the way through Ollama and an MCP server with OpenAI-compatible endpoints showing what transformations occur where.
LLMs work using huge amounts of matrix multiplication.
Floating point multiplication is non-associative:
a = 0.1, b = 0.2, c = 0.3
a * (b * c) = 0.006
(a * b) * c = 0.006000000000000001
Almost all serious LLMs are deployed across multiple GPUs and have operations executed in batches for efficiency.
As such, the order in which those multiplications are run depends on all sorts of factors. There are no guarantees of operation order, which means non-associative floating point operations play a role in the final result.
This means that, in practice, most deployed LLMs are non-deterministic even with a fixed seed.
That's why vendors don't offer seed parameters accompanied by a promise that it will result in deterministic results - because that's a promise they cannot keep.
> Developers can now specify seed parameter in the Chat Completion request to receive (mostly) consistent outputs. [...] There is a small chance that responses differ even when request parameters and system_fingerprint match, due to the inherent non-determinism of our models.
Theorizing about why that is: Could it be possible they can't do deterministic inference and batching at the same time, so the reason we see them avoiding that is because that'd require them to stop batching which would shoot up costs?
> The non-determinism at temperature zero, we guess, is caused by floating point errors during forward propagation. Possibly the “not knowing what to do” leads to maximum uncertainty, so that logits for multiple completions are maximally close and hence these errors (which, despite a lack of documentation, GPT insiders inform us are a known, but rare, phenomenon) are more reliably produced.
I, naively (an uninformed guess), considered the non-determinism (multiple results possible, even with temperature=0 and fixed seed) stemming from floating point rounding errors propagating through the calculations.
How wrong am I ?
Fielding won the war precisely because he was intellectually incoherent and mostly wrong. It's the "worse is better" of the 21st century.
RPC systems were notoriously unergonomic and at best marginally successful. See Sun RPC, RMI, DCOM, CORBA, XML-RPC, SOAP, Protocol Buffers, etc.
People say it is not RPC but all the time we write some function in Javascript like
const getItem = async (itemId) => { ... }
which does a
GET /item/{item_id}
and on the backend we have a function that looks like
Item getItem(String itemId) { ... }
with some annotation that explains how to map the URL to an item call. So it is RPC, but instead of a highly complex system that is intellectually coherent but awkward and makes developers puke, we have a system that's more manual than it could be but has a lot of slack and leaves developers feeling like they're in control. 80% of what's wrong with it is that people won't just use ISO 8601 dates.
The irony is that it's the budget cards which have fewer PCIe links. That RTX 5060 Ti 8GB only has 8 lanes, but could really benefit from having 16 lanes. An RTX 5090 with 32GB of VRAM? It has 16 lanes, but would do just fine with 8...
Renewables can absolutely provide and consume reactive power through the use of power electronics equipment incorporated into the inverters. All modern utility-scale wind and solar power plants use this type of equipment, and most interconnection requirements in the US require significant reactive power capability as part of the interconnection process. This may require additional reactive power capabilities such as capacitor banks, DVAR units, or reactors at the project substation.
This means if grid voltage drops they will provide VARs to increase voltage, and vice versa. They can similarly react to shifts in grid frequency. Only after they reach certain power factor, voltage, or frequency limits will they disconnect from the grid, similar to other generators.
If in spain renewable energy projects were somehow exempted from needing to provide a wide range of power factor capability then yes this would be a rules problem.
They charge more for inductive loads because of low power factors and significant VAR consumption yes, though you can also do power factor correction on induction motors. And though most wind turbines use induction generators, they are designed to be able to supply power at unity power factor and through a large power factor range.
Also, islanding means operating as a standalone grid, which does not happen when the generators shut down, and systems are designed to shut down after a certain number of cycles without grid. This does not sound like islanding at all.
I'm a minimalist in this regard, and I really believe that a website should only be as complex as it needs to be. If your website requires fast Internet because it's providing some really amazing service that takes advantage of those speeds, then go for it. If it's just a site to provide basic information but it loads a bunch of high-res images and videos and lengthy javascript/css files, then you should consider trimming the fat and making it smaller.
Personally I always test my website on a variety of devices, including an old PC running Windows XP, a Mac from 2011 running High Sierra, an Android phone from 2016, and a Linux machine using Lynx text browser, and I test loading the site on a connection throttled to 128kbps. It doesn't have to run perfectly on all these devices, but my criterion is that it's at least usable.
At one of my previous jobs, we designed a whole API to be slightly more contrived but requiring only one round-trip for all key data, to address the iffy internet connectivity most of our users had. The frontend also did a lot of background loading to hide the latency when scrolling.
It's really eye-opening to set up something like toxiproxy, configure bandwidth limitations, latency variability, and packet loss in it, and run your app, or your site, or your API endpoints over it. You notice all kinds of UI freezing, lack of placeholders, gratuitously large images, lack of / inadequate configuration of retries, etc.
A whole lot of chemistry process is just X dissolves in Y but not in Z, and using that in order to separate and purify.
In this case metal oxides dissolve in glass (sand, which is a silicon oxide, mostly) but gold doesn't A) oxidize under reasonable conditions or B) dissolve in the glass. Sand or glass waste is melted, the not gold dissolves into the molten glass.
The kids are not even trying to do either. They are already gravitating towards multidisciplinary teams, cause unlike past generations they are dealing with a rate of change at a totally different level. In such an environment, people get to see their own limitations much faster no matter the quality of their training and they end up having to rely on others.
The big challenge is getting very different people with ever growing different skillsets and interests to coordinate, stay in sync and row in one direction.
When the work is equal to the knowledge and judgment of the painter, it is a bad sign; and when it surpasses the judgment, it is still worse, as is the case with those who wonder at having succeeded so well. But when the judgment surpasses the work, it is a perfectly good sign ; and the young painter who possesses that rare disposition, will, no doubt, arrive at great perfection. He will produce few works, but they will be such as to fix the admiration of every beholder.
Leonardo da Vinci, "A Treatise on Painting.", p. 225
I think these are key thoughts worth considering going forward:
> Code is not an asset, it's a liability.
> Every line must be maintained, debugged, secured, and eventually replaced. The real asset is the business capability that code enables.
> The skill that survives and thrives isn't writing code. It's architecting systems. And that's the one thing AI can't do.
> Any time you use an LLM with tools that might be exposed to malicious instructions from attackers (e.g. reading issues in a public repo, looking in your email inbox etc) you need to assume that an attacker could trigger ANY of the tools available to the LLM.
I think we need to go a step further: an LLM should always be treated as a potential adversary in its own right and sandboxed accordingly. It's even worse than a library of deterministic code pulled from a registry (which are already dangerous), it's a non-deterministic statistical machine trained on the contents of the entire internet whose behavior even its creators have been unable to fully explain and predict. See Claude 4 and its drive to report unethical behavior.
In your trifecta, exposure to malicious instructions should be treated as a given for any model of any kind just by virtue of the unknown training data, which leaves only one relevant question: can a malicious actor screw you over given the tools you've provided this model?
Access to private data and ability to exfiltrate is definitely a lethal combination, but so his ability to execute untrusted code, among other things. From a security perspective agentic AI turns each of our machines into a Codepen instance, with all the security concerns that entails.
This is why capitalists dislike unions so much, becasue they know this. Together we are stronger.