The harness is the software that runs the model, and, in the case of agents, implements the agentic loop, and provides capabilities external to the model (either directly or by interfacing with other systems.)
You need a harness for ANY use of a model because otherwise you just have a bunch of numbers that are the weights and no actual software that does anything with them, and you can’t have agents without a harness that provides the capabilities that define agents.
> I am no writer but to me it seems like writing is one of the easiest tasks with barely any logic or reasoning
Virtually all logic or reasoning is, in one way or another, part of the support for writing. It’s what separates actual writing from generating nonsense that happens to fit grammar rules.
The specific details depend on the domain, of course, but I can’t see how anyone familiar with the output of writing can think that there is little logic or reasoning in doing it well.
Unknown quality is the absence of information, you can’t have a signal for it.
If you are using “unknown” as a euphemism for “poor”, which is the most obvious thing that makes the signal meaningful in the direction that seems apparent from context, why?.
> I disagree. It is not discrimination between 2 people to charge one $200 for a 2 hour private lesson and the other $100 for a 1 hour private lesson.
Sure it is discrimination on the basis of duration sought, not discrimination on the basis of personal characteristics, and while it is discrimination, its fair to characterize it as bot discrimination between people, since different people making the same choice would be charged the same. Its obviously, in any case, not discrimination anyone would object to, but objectionability is a matter of the basis of discrimination, not the fact of discrimination.
Not always shell—Python, often a subset, is common—but a single sandboxed coding-running tool as the way to run all of the other things that would otherwise be their own top-level tools themselves seems an increasingly common approach.
> Charging people a price relative to the cost of a service is not discrimination.
Yes, it is discrimination. Treating people differently on any basis is discrimination on that basis.
It may or may not be illegal discrimination depending on the specifics and the governing law, and it may or may not be unjust discrimination depending on the specifics and the applicable moral framework, but all discrimination means is differential treatment on some criteria.
> Abusive controlling partners rarely are long-term timeframe rational, because being long-term timeframe rational is at odds with being an abusive controlling partner.
Humans are rarely “long-term timeframe rational” (rational choice theory is very much a known-false model of human behavior on both the individual and aggregate level), but there is nothing about abusively exploiting an initially trusting and eventually dependent telationship that it is inconsistent with long-term rationality.
> Same outcome, different implementation, instead of compiler plugins or bytecode rewrite, it gets patched via the languages dynamism.
While the robust dynamism of Ruby (or Python) can allow it to be used in a patching-like way, the only dynamism that is essential to the implementation is the dynamism of runtime polymorphism (common in statically-typed, AOT-compiled OO languages.) The language feature it relies on that C++ and Java don’t have is multiple inheritance that gets resolves to a defined order you can call back up from a concrete method with “super”. There is no reason a static OOP language that supports runtime polymorphism couldn’t support this; it wouldn’t even need to add overhead to compiled classes that don’t make direct super calls from their own methods.
I can't imagine that model lifetimes will ever justify using model-specific ASICS for public serving (maybe something like serving fixed certified AI models in a vehicle or robot) over more generic GPUs/NPUs until after the AI bubble pops.
> Also one would say monkey patching on Python and Ruby frameworks is another way to do AOP.
IIRC—and it may have changed in the several years since I made use of it for this, but I don’t see why it would—the standard way to do AOP in Ruby is leveraging modules-as-mixins which are a core language feature, monkey patching is unnecessary (but since classes in Ruby are open, modules-as-mixins can be used to monkey patch classes provided by someone else just as easily as being done at class definition time.)
Aspect-oriented programming has annoying implementation in languages like Java which don’t natively support the right abstractions and where you are fighting the language to do it.
Its kind of unfortunate that those are also the languages also which do the most to shape people’s understanding of AOP.
> That's the classic question of when the state should step in to correct bad parenting, and when it should leave well enough alone
Well, no, this is the slightly different classic question of whether, in order to prevent (or, more realistically, on the pretext of preventing) bad parenting, the state should impose a universal totalotarian papers-please approach to basic life activities which requires everyone to prove identity to refute the assumption that they are a child so that it may impose its own (also bad, because totally one-size-fits-all and corcumstance blind) parenting on every child, without regard to the imposition on the liberty of every adult necessary to do so.
Stepping in to correct bad parenting happens when there is a system to identify bad parenting requiring correction and intervene only where such is identified, not when you are imposing a universal regime on everyone on the justification that without it, some will act as bad parents. While both are approaches frequently proposed, they are different and should be distinguished.
That might seem an attractive intuition based on observing damage and having an understandable intuition of how damage should relate to energy, its wrong.
Energy is force × distance. For gravitational potential (and, equivalently, the kinetic energy acquired when that is tranformed into kinetic energy by a fall), the force is the force of gravitational attraction between the body and the object it is falling toward (the object’s weight) and the distance is the distance of the (actual or potential, as appropriate) fall.
The nonlinearity you observe in damage is because damage is a complicated outcome that depends on a lot more than aggregate energy involved.
> Any superintelligence operating under a consistent moral framework will decide to extinguish humanity with as little ecological damage as possible, because humans cannot coexist with other life on this planet.
There are plenty of internally consistent moral frameworks which would not favor this action even if the premise were true (and that premise seems at best unjustified and at least overstated.)
> Prohibition and The War on Drugs in the US are two good examples of something that had altruistic origins
The War on Drugs had the purpose (not just in its origin but in its perpetuation) of inflicting harm on elite-disfavored subsets of the population that could not be openly targeted for Constitutional reasons, which is about as far from an altruistic reason as it possible to get.
The word “openly” in my post there for a reason; the commercial models are not openly distilled from competitors: many open source models have in their model documentation that distillation was done from a dataset drawn from specific other models, including commercial models.
That distillation might be inferred from the behavior of commercial models is not the same as them openly doing it.
> > Why would you ask for a self-reported, unverifiable test score that could be decades old at this point?
> It's true that self-reported scores are not the most accurate, but if I were applying for a job I would report honestly, on the assumption that they could easily request for the scores to be sent by the College Board.
No, they couldn't, except by going through you (the College Board doesn't take third-party score requests.) You might be able to request that if they are recent enough, but not if they are literally decades old (well, not if they are ~21 years old or older.)
The dot-com bust produced an EXTREMELY mild recession (so mild that it was often misattributed to 9/11, which occurred when it was almost over.
OTOH, there was a lot of pain iny the subsequent expansion leading up to the 2008 , but that was all the fault of fiscal (eepecially tax) policy of the Bush Administration and thei Congressional allies, not Fed monetary policy. While Greenspan clearly ideologically supported the people doing that, it wasn't him and the Fed causing the problems.
Not revealing actual thinking traces prevents mdoel distillation on yhe actual output (thinking traces are a key part of the output) which makes it harder for conpetitors to catch up (a moat).
Being currently in the lead in a category is not a moat,a moat is whatever creates a barrier to competitors catching up when you are in the lead. Merely being in the lead is not a moat except in a market with strong network externalities.
You need a harness for ANY use of a model because otherwise you just have a bunch of numbers that are the weights and no actual software that does anything with them, and you can’t have agents without a harness that provides the capabilities that define agents.