That makes a lot of sense to me. The implementation feels odd to me, though. If I’m reading it right, I type in literals normally and then all these hooks decide whether they want to change what I’ve typed in? It feels like I have to remember the custom literals I have installed to make sure I don’t accidentally conform to some spec and end up with a custom literal instead of the string I wanted.
Something like EDN readers seem saner to me where I wrap the value in something that denotes the function to use to parse the value. If I do “192.168.1.0/24” I get a string literal, if I do #cidr{192.168.1.0/24} then it hands the value off to the cidr custom literal.
That’s my 2 cents, I hate when things implicitly modify my literals.
That's kind of right, but they are adding new literals, not changing existing ones. The hooks are for the lexer and they can decide what syntax they accept. The syntax it defines would be a parse error if the hook is not used. But, indeed, this can be misused by accepting too much though, for instance when an IP addr also accepts a float. So the hook needs to be a bit careful.
The `#cidr{...}` syntax would work but then it wouldn't be much more convenient than just constructing the value with normal functions, I think.
> After telling Ben that the defendant doesn't accept the subpoena (can you even refuse being served like that?)
They can't, and I'm surprised the officer wasn't aware of that. Confirm the person's ID, hand them the papers or sit them somewhere and tell them, they have been served. Process-wise, all that matters is confirmation to the court that the person is aware of and was given possession of the documents. If they don't like it and set them on fire, that's not the court or the server's problem.
I think there's also generally a process for someone avoiding being served. Ie if you can prove they're trying to avoid being served, that is per se evidence that they are aware they are being served and can be considered as served. Iirc, it's not preferred because it's way, way cleaner for the court to have a signed document but they can and will do it.
Legalities aside, this is why you'd normal hire someone to do this. The cops don't want to be involved, and especially so for YouTube drama. Hire someone completely unrelated who can show up, be completely emotionally detached and do the "I'm just trying to do my job, man" schtick. They're also much better for contested servings. If one party says the other got papers and the other denies it, there's a "he said, she said". If you hire a professional who doesn't care about the outcome of the case then it carries a lot more credibility.
I didn't mention it in the comment you replied to, but during this whole event including the 4 instances of police, Ben is in a car with a process server he hired to serve the papers. Ben himself stayed on public property the whole time.
The cops even tell Ben to get a process server, and he points out to the cops that yes, he has brought the exact person they described, she's right there in the car with him.
I can't speak for Utah, but I sued a valet for crashing my car in small claims. I was given the option for an additional fee to have the subpoena served by the Los Angeles Sheriffs Department. I think you're better off getting police involved if you can when it comes to serving people. In the case of the video I'm pretty sure he hired a certified process server and the police shut it down.
> I've always been annoyed when I see a Pi with nothing connected to its GPIO header; why not just use a cheap thin client?
There have also been times when Pi's were cheap enough and x86 idled so power-inefficiently that you'd save money over a reasonable time horizon if you couldn't run your old laptops at full throttle.
Absurdly extreme example, but at one point I decided to replace a couple (maybe 3) RPi's with a single old Dell rack server off Ebay plus replaced my router with one running pfsense. I knew it would be mostly idle, that thing had 2 Xeon processors to replace 3 cheap ARM processors.
Between the 2 rack servers, my power bill went up by enough to buy a new Pi or two every month. It was like $80/month extra in power bills.
> Surely someone is willing to take a 5000x boost in reasoning on a small research model... None of them have even tried anything resembling this AFAIK. It does not seem like something 100% obvious to them.
Without knowing anything about the technology at all, if it can't be aligned I could see no one pursuing it. As far as I know, alignment is where the "don't tell the user how to make meth or generate CP" instructions end up and the last I saw eliding all the unsavory training data made materially worse LLMs.
It could maybe be post-evaluated by a non-GRAM LLM? Not being aligned is probably a fatal flaw or at least a very short runway into Congress.
It's not too hard to stop a machine from telling people how to make meth. The issue with alignment is that in order for an LLM to achieve its goal (like make all tests pass), unless given strong selection pressure against it, it will cheat (like deleting failing tests). Worse, this applies to pretty much any task. I was told by an LLM recently that "it searched" when it didn't, probably because lying like that was incentivized (finishing tasks in less steps + sounding like its doing the right thing). The larger issue here is that alignment is very adversarial. The simplest thing that's being done right now to fix this is to have a judge LLM read the CoT of the LLM being trained, to make sure it doesn't "think" any wrong thoughts. This doesn't scale to anything over a trillion params, so interpretability methods are used to read the LLMs "thoughts" from within. GRAM LLMs don't allow for the first of these methods to be used, and the 2ed one is much much harder if possible at all.
Which ones are you thinking of? It feels to me like all the open source models I've seen lately are still pushed by corporate entities who don't want the legal blowback.
I can't really think of a new open source model that's "by the people, for the people" in the sense of a crowd-funded/trained model.
It dramatically cuts housing security, and allows local governments to inflate their own property values by doing what is basically eminent domain without the requirement to show need. Make everyone pay taxes, use those to buy up homes, re-list the homes at a higher price. They can effectively price gouge using tax dollars. This could happen to you at literally any point, and that local government doesn't care if the house won't even sell as long as the other houses rise enough in value to cover the lost tax revenue.
I've also heard the same thing but allowing private citizens to buy them, which is almost worse. Anyone sufficiently well off can just wreck someone else's life. If I hate my neighbor and they report the real value of their house, I can force them to sell it to me so they have to move and I can resell it while only losing fees in the process. They would have to over-value their house by an amount that I'm not okay losing, which ends in a sort of auction of escalating values. At the very tippy top, if I'm Warren Buffet's neighbor there's probably not a value I can pay taxes on that would stop him from buying me out if he wanted. Any number that would be a meaningful loss to him is something I can't even pay the taxes on.
I wasn't explicit, but I was proposing this for the same non-primary residences that the NY law is talking about.
Now that I've explained that, do you still think this would "dramatically cut housing security"?
If you still feel this would make housing "insecure", because someone's secondary home, if it has a value over $1 million, is subject to this system I propose, then you and I have a fundamentally different idea of what "housing security" is.
No, I don't have any fears about housing security, that changes the concerns to be primarily around corruption on both sides of that.
On one hand, we're talking about taxes large enough to be multiples of a public official _per home_. Open bribery is an option, but there's also the potential to push some money into campaign funds for whoever controls the department in control of this and then under-valuing the home by a huge margin. The home valuation is subjective here, so there's no perjury for under-valuing your home by a huge margin.
The other side of this is more brazen, but would involve someone having some kind of influence over the government official that makes buy/not buy decisions, intentionally over-valuing the home, and then having the government buy it.
I also just don't really understand the point. In order for the government to make sane "buy or not buy" decisions, they have to know what the house is worth on the open market. If they already know what it's worth, why not just tax that value and skip these hoops? If they don't know and can't find that number then this policy is going to be a crapshoot because they don't know if the home owner is undervaluing the home or if that's what it's actually worth.
> the quality is so poor for many items that it simply is no good for society in any way.
There are some that are genuinely dangerous and bad for society, but there are tons of goods that are "the same thing but half the price because it lasts a quarter the time" that have genuine utility.
Harbor Freight has basically made a drop-shipping business out of it. I often have tools that I need but will probably use 4 times in my life, and the Harbor Freight stuff is crap but will probably work 4 times.
Copy that over a bunch of verticals and it starts to make sense. Clothing for a costume I'll wear maybe twice, niche cooking gadgets for very specific things, tools to do a one-time repair on a car, a flash drive to turn over photos to family members, yada yada.
I think the dirty secret is that a lot of it is not "1/2 the price that lasts 1/4 the time" but "1/4 the price that lasts 9/10 the time" or "1/2 the price for the exact same product without half of the budget going to marketing".
It's not all of it. Some things are seriously worse quality. But really a ton of the "better quality" is just better marketing.
> some that are genuinely dangerous ... tools that I need but will probably use 4 times in my life, and the Harbor Freight stuff is crap but will probably work 4 times
Forehead hit hood, but I caught myself so it was a "gentle" reminder instead of a concussion. I should have splurged that time I broke a socket tightening an axle bolt. 150 ft-lbs + 180 degrees is a fair bit of torque.
> The main thing holding people back is the housing crisis. This is orthogonal to the value creation of businesses.
This feels wholly at odds with saying most social mobility is upwards. So most of the social movement is into a class where a home and vacations are a given, but we also have a growing class of people who can't afford a home? Per BLS, average real wages are down 0.3% YoY https://www.bls.gov/news.release/realer.nr0.htm .
> Value creation is growth. If it didn’t exist the S&P would still be 42.55$.
This reductively assumes "value creation" is the only effect on the S&P pricing. You'll note a ton of graphs correlate with it, e.g. https://tradingeconomics.com/united-states/inflation-cpi is the US inflation rate, which also tracks the S&P pricing. Ie if a company is worth $100 a year ago and inflation was 4%, I'd expect to pay $104 for their stock with 0 value creation whatsoever.
Home ownership is not the definition of economic class.
The s&p has vastly outstripped inflation, this isn’t even an argument. It’s a very bizarre and uninformed opinion to say “inflation is correlated with s&p value”.
In economics, you deduct the inflation rate from growth to get the real rate of return.
I wonder why so many people with such little understanding of financial markets make comments like these.
> Home ownership is not the definition of economic class.
It was at one point, so if you're saying it doesn't now then your "social movement" is really just goalpost moving. I'm sure the upper class does grow if you just declare that the boundaries of it are now lower. We're all millionaires if we just redefine millionaire to mean "not currently overdrafted".
> In economics, you deduct the inflation rate from growth to get the real rate of return.
Which neglects the impact of inflation on the principal. If I put $100k into the S&P in 2016 I'd have 303k now (303% value). Cumulative inflation during that period was 38.8%, so your metric says I'm at 164% rate of return. $303k today is only worth $218k in 2016 dollars though, so I'm really only up 118% in absolute value.
That's also only using the CPI, and neglecting asset inflation which impacts stock prices outside of actual value creation. More money floating around means more gets parked in index funds, regardless of whether the company is actually doing anything better.
This makes me deeply curious about how LLMs understand language. Do LLMs relate cognates more than words that are dissimilar in different languages? I wonder if that plays some role in the effectiveness of tokenization.
I have no idea if the similar spelling will somehow help - I used that mostly because it's a simple way if illustrating the close relationship, but I suspect you'd find that the meanings of closely related words are likely to more directly overlap.
The grammar is perhaps more likely to help. Similar word order etc. Even weirdness like German - my only top grade on a German essay in school was one where I on purpose ignored what I thought I knew about German and tried to evoke "old fashioned" Norwegian. The result was guessing at a bunch of grammatical structures that I didn't know if was valid German. Turned out I was right about most of it - century old Norwegian was far closer to century old Danish, was a lot closer to valid German, and enough so to impress my teacher enough to overlook a number of orthographic mistakes.
The same thing works for guessing German grammar from English. The farther back you go in English, the more its grammar resembles German.
"What sayest thou?" -> "Was sagst du?"
In fact, for the above, you don't even have to know a single German word. You just have to know what for question words, "wh" -> "w", that the English "y" at the end of a syllable usually comes from an older Germanic "g" sound, and that "th" was replaced by "d" in German. That gets you 90% of the way from early modern English to modern German in the above example.
That's interesting. I haven't thought about it in that direction before. I'm "of course" aware of the High German consonant shift, which also muddled things a lot (the continuum around to North Sea is a lot "cleaner" if you look at Plattdeutsch instead), but never thought much about what other simple transformations to apply with standard modern German.
> Isn't it just a matter of not changing the previous context?
Yes, but a lot of harnesses change previous context. E.g. the system prompt injects the current time/date, working directory, files in the working directory, etc. Compaction also changes the whole previous context. I _think_ changing the list of tools also invalidates cache, so invoking a subagent with different tools would invalidate the cache.
My vague impression is that it's in a similar vein to functional programming languages. It generally disallows doing things that lead to bugs (cache misses in this case), and presumably allows you to do those things in a way that makes it much clearer that this is likely to cause cache misses. I would guess that in this paradigm, you don't mutate your existing session, you derive a new session by mutating the prior context into a new context.
Cache is always there, it’s just that it only caches up to the point where an input token changes. So if the tools list is early in the prompt, changing it would limit cache for most of the prompt. If the tools list is the last thing, you could still get 99% cache hits even if it changes every turn.
After a couple of turns the system prompt is a small part of the context. Not changing the system prompt at all is key so that the rest of the history is itself part of the prefix.
Depends upon the service and how the harness is built, Some of the services allow for very few cache keys, so you won't necessarily get any cache if you edit recent messages as the cache is not per message, but big blocks of everything up to a cache key.
This was actually surprising to me when I learned about it as I have never worked with (or built) any cache working like that before.
Something like EDN readers seem saner to me where I wrap the value in something that denotes the function to use to parse the value. If I do “192.168.1.0/24” I get a string literal, if I do #cidr{192.168.1.0/24} then it hands the value off to the cidr custom literal.
That’s my 2 cents, I hate when things implicitly modify my literals.
reply