Rendered at 06:56:58 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
socketcluster 9 hours ago [-]
It is true that they don't tend to give simple answers.
But it reveals something about human biases that we tend to seek out simple answers to complex questions. We tend to think that because a question is short, that it's simple. We forgot how many details are involved and how much is actually going on underneath. Reality itself is often annoyingly complex.
All living creatures, including humans, are conditioned to simplify extremely complex things into simple concepts which we can react to. We only understand things to the extent that we believe they affect us.
goodmythical 8 hours ago [-]
That's just a caloric efficiency thing, isn't it?
Far easier to simply spread peanut butter and jelly on bread than to consider how the peanuts were harvested, how many bugs were in the fruits, how close the wheat was to molding, how much sugar is in the bread, how many years it tooks us to perfect such tasty and long lasting ingredients, what potenitally substandard conditions the labor involved endured...etc, every time one wants a sandwhich.
And then a multiple of that easier as well in conversing about a peanut butter and jelly sandwhich as few if any of the parties will think about any of the above.
We see the same in computers. A web dev can make it quite far without once considering the OSI model. A game dev can complete an entire career without ever worrying about the underlying bytecode.
There are situational benefits to diving deeper, of course, but I don't imagine we'd make it very far if we had to reinvent the universe for every word in every conversation or chain of thought.
grayclhn 1 hours ago [-]
>> But it reveals something about human biases that we tend to seek out simple answers to complex questions. We tend to think that because a question is short, that it's simple. We forgot how many details are involved and how much is actually going on underneath. Reality itself is often annoyingly complex.
> That's just a caloric efficiency thing, isn't it?
perfect, no notes
jvanderbot 5 hours ago [-]
Calorie, bit also speed of decision and easier transmission to peers and kiddos.
A quick rhyme encoding a useful heuristic? They'll remember that for generations!
proc0 7 hours ago [-]
Simple answers are the result of a problem being solved all the way through. It abstracts away all the details and exposes the essential part of the solution. This is the compression perspective of intelligence. We want things as compressed as possible because it's how intelligence more or less works.
Shorn 2 hours ago [-]
Whitehead said it in 1911:
> “Civilization advances by extending the number of important operations which we can perform without thinking about them.”
He followed it with:
> “Operations of thought are like cavalry charges in a battle — they are strictly limited in number, they require fresh horses, and must only be made at decisive moments.”
b3ing 5 hours ago [-]
Problem is tech is full of PMs and other managers that don’t know technology. Things that could be algorithms, functions, programs, regex filters, etc are now LLMs because they don’t know anything else to solve problems
cookiengineer 5 hours ago [-]
> Problem is tech is full of PMs and other managers that don’t know technology. Things that could be algorithms, functions, programs, regex filters, etc are now LLMs because they don’t know anything else to solve problems
Is this the right approach though?
Computers should be usable by humans, not the other way around. How CPUs are built, how assembler came to be, how programming languages came to be... ultimately, they were all just compromises along the way to make computers more usable by humans.
With LLMs, however, the threshold barrier for being able to modify and use a computer effectively is now zero, which in my opinion is a good thing.
I can recommend watching "the mother of all demos" talk from 1968 [1] which basically established the design patterns for modern computers. That's how old the concept of desktop environments is. There hasn't been any meaningful invention eversince, apart from touchscreens/smartphones, and now LLMs as a generalized interface design.
Well LLMs are more expensive and take longer to setup. If non tech people want to build and leverage APIs and so forth for something they could have done in 1 hr or less of coding I guess that’s up to them. But I’ve seen them planning thousands of dollars costs and a team of people working on something simple
whycombinetor 2 hours ago [-]
This year I listened to someone suggest using an LLM call to go from the UI field contents of a web form to the JSON to send over the wire. Like, a .map statement.
perrygeo 2 hours ago [-]
This year I listened to someone use an LLM call to sort a list of dictionaries by key.
The LLM did indeed sort the list - after several seconds.
My editor can sort the list in single-digit milliseconds.
Do you really aspire to be the 0.001x developer? Weird flex, bro. It's just embarrassing.
There are operations best done locally. Learn your tools. Outsourcing these task simply because you're unaware is not something to be proud of.
jamesfinlayson 22 minutes ago [-]
My old boss proudly told all how he'd used Claude to create three tickets in Jira. He said it took 10 minutes and normally it would take him 2 minutes, but he was still thrilled.
lhk931122 2 hours ago [-]
If this direction works out, I think service companies like Anthropic or OpenAI would provide it. If there is a lot of demand. Until then you have to set it up yourself locally, which is a hassle. Definitely, for components that act as sub agents in the agents I use personally, this direction actually looks useful for getting answers that rely even a little less on randomness and using them.
simianwords 2 hours ago [-]
OpenAI already provides Luna based decision model
throwawayffffas 7 hours ago [-]
Using these decision models as a model router sounds really bad to me the models will most likely be newer than the decision model and thus the decision model will have no information about the capabilities of the model other than the short blurb you feed it.
nostrebored 2 hours ago [-]
I am so skeptical about these decision model rerankers. Training a reranker has never been easier, and many of them already return calibrated scores.
Decision models are often larger, slower, non colocated… it just seems like a strange tradeoff to make.
dzink 9 hours ago [-]
PostHog should not be using them on customer data, period.
jakeydus 9 hours ago [-]
Every time one of these consumer data brokers/gatherers shows some insight into how they gather the data it makes me feel sick. I know they're doing it, I know it's granular, it's still gross.
nixosbestos 8 hours ago [-]
Gathering it and pumping it all through Anthropic and OpenAI data collectors.
goosejuice 9 hours ago [-]
I mean look at what they sell. Surveillanceware.
dominotw 7 hours ago [-]
postslop
nandanadileep29 52 minutes ago [-]
very true, many LLM usecases can be replaced with decision models like these, its wayy cheaper and faster.
purusa0x6c 1 hours ago [-]
exactly.. ppl are still using LLMs agents for simple direct workflows..
As I understand it (and count that I will be proven wrong if not), Strands Decider is an LLM, just a discriminative example instead of a generative one.
fauria 6 hours ago [-]
Strands Decider is an open source decision model like Jev, they serve the same purpose.
Hearing about use cases for decision models and seeing how they are being explored is quite interesting, but I think the title of this post is poorly chosen.
aabdi 9 hours ago [-]
Contrarian take.
I think everyone should use LLMs and the most expensive possible, unless they know how to run evals (both on line and off line).
Yes its expensive, ugly, etc.
But...
I've seen enough IRL that people just create really bad evals that do not reflect real world. Then the product is just bad. Outright broken for the p90 customer.
People don't know statistics. Offline Evals are hard. Sampling the entire population distribution is hard. Sampling the fat tail is harder. Online evals are big ops projects that are giant balls of death.
abhgh 8 hours ago [-]
Maybe I should write up a post about this, but about evals ... I have feelings.
Yes you're right that writing evals is hard. And it comes down to the simple fact that the thing that you're testing - a LLM-driven feature that presumably suffers from variance and inadequate accuracy - and your measuring device - an LLM judge - suffer from the fundamental problems. But the lure of natural language is such that instead of seeing this pipeline of LLM-driven things as a path of cascading errors, it is easy to view it as one set of problems correcting for another set.
In this world, the LLM-driven feature is a probabilistic entity, and the judge layered on top of that is another probabilistic entity. The real work is to look at the judge's response and infer the performance of the feature. Good evals make the inference step "pass-through", i.e., the judge outcomes have a 1:1 mapping with feature outcomes. But this kind of reliability often doesn't naturally appear. I think of reliability as being composed of three attributes: (a) consistency: do we get the same result for an eval run on the same llm with the same input (b) inter-llm agreement: do outcomes agree over multiple trials of running an eval over multiple llms with the same input (c) accuracy: does the eval agree with human judgement. You need extra work to guarantee these properties. And even when you can, you won't get 100% reliability, but say high 90s, and you still to make a judgement call as to call a feature "functional".
aabdi 7 hours ago [-]
Two things, my problem isn't with the probabilistic nature, its:
1. with the quality of the data that customers use to eval.
2. with the lack of rigor/ops/testing around accuracy for deployment and regression.
I'll give you examples for failures IRL:
Offline:
- sample distribution for test data didn't have p99 customers in it, failed in production when it launched for those customers.
- test data sets were all synthetic, launch to real customers fails because data set was not representative of real.
Online:
- yeah we used LLM as a judge in prod/online eval, but the judge sampling failed to measured against our japanese customers, our bilingual customers, customers with media devices etc.
- the LLM as a judge was used, but there was no measure of human accept rate measures/alarms for various use cases (lightbulbs turning off, music playing, etc) so the lightbulbs started failing.
- there was no consistent human sampling of online data, so for example samsung TVs stopped working, but no one caught it until 3 months later when some VP was watching his TV and it broke and was accelerated despite it affecting 2-3% of customers.
dieselgate 8 hours ago [-]
Can you elaborate on what you mean by Eval? All I can think of is eval() in js...
I hope you agree people should use the "right tool for the job" meaning a calculator for arithmetic rather than an LLM for everything.
aabdi 7 hours ago [-]
Evals are tests of model accuracy. For example here are 100 people's sample houses, go test if the agent can turn off the lightbulbs correctly.
Right job sure. But, If you're replacing the LLM with a jev like as in the article, or switching with a smaller classifier, then you better have good data/evals otherwise you're reducing accuracy without you even knowing.
dominotw 7 hours ago [-]
its basically
assertEquals(x,y) where function is calling llm to check equals
mrbungie 7 hours ago [-]
If you are not getting feedback from those p90 then your bad evals are the least of your problems.
Also, unless those p90 customers are bringing a big portion of your income (again, if that happens you'd KNOW that your product is broken for them) I'd not pay for the most expensive LLM to "avoid doing evals".
aabdi 6 hours ago [-]
>If you are not getting feedback from those p90 then your bad evals are the least of your problems.
the feedback is part of your online evals... if you have bad feedback you have bad evals
> Also, unless those p90 customers are bringing a big portion of your income (again, if that happens you'd KNOW that your product is broken for them) I'd not pay for the most expensive LLM to "avoid doing evals".
they are but that's besides the point.
Two ways:
- logically ROI of accuracy degradation as compared to cost has to be measured, sure.
- emotionally, morally sub 10% groups in our customer distribution include such things as disabled people, minorities, and the elderly. In my view you should love all your customers.
sick_of_slop 7 hours ago [-]
[dead]
HacklesRaised 2 hours ago [-]
almost nothing in fact...
saejox 8 hours ago [-]
I have been using structured json outputs since gpt 4 days. Never understood this jev hype
acedTrex 8 hours ago [-]
i think the main deal is speed right?
triMichael 6 hours ago [-]
No, the main deal is that it CANNOT answer with anything outside the given structured options. LLMs can go off the rails, however small a chance it is. The architecture behind JEV and the like doesn't let it do that.
j16sdiz 5 hours ago [-]
> structured json outputs
vibebingo 6 hours ago [-]
> What is the current conservation status of hedgehogs?
A harnessed LLM can look this up using tools.
Maybe LLM at least ti prepare context for decision model.
I guess if using a decision model context prep is key.
dieselgate 9 hours ago [-]
Completely unrelated to the article itself but I landed on the posthog.com homepage after viewing their open positions and the landing page is cool. It's sort of a desktop UI and different from what I usually see, was expecting some other HR domain to come up.
jakeydus 9 hours ago [-]
"Not everything needs an LLM! Use a decision model!"
And what is a decision model exactly? An LLM with instructions to limit responses?
Huh.
orbital-decay 8 hours ago [-]
Sure it's an LLM, it's just neither generative nor instruction-tuned, like BERT (also an LLM), which gave them a reason to market it as not being one. Though modern "LLMs" are also modeling a lot more than just the language so they're kind of hard to define. Not much point in being pedantic, let them call anything anything and see what sticks.
amelius 8 hours ago [-]
Sounds like a yes/no question.
tchalla 6 hours ago [-]
Not everything needs a "decision model" with a "system one" API either but here we are.
Rapzid 8 hours ago [-]
The approach of using decision models to gate more expensive work.. I mean, yeah.
But using protecting PII as an example use case is a bit bonkers?
jscd 7 hours ago [-]
I continue to be astonished that a real company exists called "post hog"
Terr_ 7 hours ago [-]
Post hog, ergo proper hog.
simianwords 3 hours ago [-]
The fact that Jev can't do reasoning means that it can only be used for very basic decisions.
Jev can't answer this for example
What is 6^8-3&2 mod 2? 0 or 1?
You might think it is a contrived example but it is illustrative. It shows that almost most questions are reasoning shaped.
11 hours ago [-]
cognitif 8 hours ago [-]
Wait, I don’t need an LLM to do arithmetic?
TacticalCoder 4 hours ago [-]
> A probability score of 0.85 still means ~150 wrong answers per 1,000 checks, so plan ahead for false positives and false negatives.
But that's only true if we believe the value of 0.85 in itself is correct!? For all we know it could be 0.75 or 0.9 no?
If were to make calls to something like JEV and know that the probability the answer is correct is, say, 0.85193847098713, wouldn't that be even more amazing than JEV and LLMs themselves!?
I'm not being facetious here: how do we know it's 0.85 and neither 0.75 nor 0.9?
And if 0.85 is "correct", why stop there? Why not 0.85193847098713?
Any explanation is very welcome.
thereader12 4 hours ago [-]
Yeah, if I understood correctly, they are using it wrong. The value is Jev's confidence which they trained for higher accuracy but it's still Jev is 85 percent sure not it's an 85% chance of being correct.
brodydung 4 hours ago [-]
[flagged]
chanhyounglee 2 hours ago [-]
[flagged]
johnh-hn 8 hours ago [-]
[dead]
shipitx 5 hours ago [-]
[flagged]
ClayCantrell 8 hours ago [-]
[dead]
pj_mukh 8 hours ago [-]
tl;dr Not everything needs an LLM, but everything needs a typesafe wrapper around an LLM.
Gotcha
acedTrex 8 hours ago [-]
Ridiculous field we are in, people seemingly have forgotten the absolute basics in under a year.
LtWorf 8 hours ago [-]
The people doing these things never knew the basics.
But it reveals something about human biases that we tend to seek out simple answers to complex questions. We tend to think that because a question is short, that it's simple. We forgot how many details are involved and how much is actually going on underneath. Reality itself is often annoyingly complex.
All living creatures, including humans, are conditioned to simplify extremely complex things into simple concepts which we can react to. We only understand things to the extent that we believe they affect us.
Far easier to simply spread peanut butter and jelly on bread than to consider how the peanuts were harvested, how many bugs were in the fruits, how close the wheat was to molding, how much sugar is in the bread, how many years it tooks us to perfect such tasty and long lasting ingredients, what potenitally substandard conditions the labor involved endured...etc, every time one wants a sandwhich.
And then a multiple of that easier as well in conversing about a peanut butter and jelly sandwhich as few if any of the parties will think about any of the above.
We see the same in computers. A web dev can make it quite far without once considering the OSI model. A game dev can complete an entire career without ever worrying about the underlying bytecode.
There are situational benefits to diving deeper, of course, but I don't imagine we'd make it very far if we had to reinvent the universe for every word in every conversation or chain of thought.
> That's just a caloric efficiency thing, isn't it?
perfect, no notes
A quick rhyme encoding a useful heuristic? They'll remember that for generations!
> “Civilization advances by extending the number of important operations which we can perform without thinking about them.”
He followed it with:
> “Operations of thought are like cavalry charges in a battle — they are strictly limited in number, they require fresh horses, and must only be made at decisive moments.”
Is this the right approach though?
Computers should be usable by humans, not the other way around. How CPUs are built, how assembler came to be, how programming languages came to be... ultimately, they were all just compromises along the way to make computers more usable by humans.
With LLMs, however, the threshold barrier for being able to modify and use a computer effectively is now zero, which in my opinion is a good thing.
I can recommend watching "the mother of all demos" talk from 1968 [1] which basically established the design patterns for modern computers. That's how old the concept of desktop environments is. There hasn't been any meaningful invention eversince, apart from touchscreens/smartphones, and now LLMs as a generalized interface design.
[1] https://www.youtube.com/watch?v=yJDv-zdhzMY
The LLM did indeed sort the list - after several seconds.
My editor can sort the list in single-digit milliseconds.
Do you really aspire to be the 0.001x developer? Weird flex, bro. It's just embarrassing.
There are operations best done locally. Learn your tools. Outsourcing these task simply because you're unaware is not something to be proud of.
Decision models are often larger, slower, non colocated… it just seems like a strange tradeoff to make.
Yes its expensive, ugly, etc. But...
I've seen enough IRL that people just create really bad evals that do not reflect real world. Then the product is just bad. Outright broken for the p90 customer.
People don't know statistics. Offline Evals are hard. Sampling the entire population distribution is hard. Sampling the fat tail is harder. Online evals are big ops projects that are giant balls of death.
In this world, the LLM-driven feature is a probabilistic entity, and the judge layered on top of that is another probabilistic entity. The real work is to look at the judge's response and infer the performance of the feature. Good evals make the inference step "pass-through", i.e., the judge outcomes have a 1:1 mapping with feature outcomes. But this kind of reliability often doesn't naturally appear. I think of reliability as being composed of three attributes: (a) consistency: do we get the same result for an eval run on the same llm with the same input (b) inter-llm agreement: do outcomes agree over multiple trials of running an eval over multiple llms with the same input (c) accuracy: does the eval agree with human judgement. You need extra work to guarantee these properties. And even when you can, you won't get 100% reliability, but say high 90s, and you still to make a judgement call as to call a feature "functional".
1. with the quality of the data that customers use to eval.
2. with the lack of rigor/ops/testing around accuracy for deployment and regression.
I'll give you examples for failures IRL:
Offline:
- sample distribution for test data didn't have p99 customers in it, failed in production when it launched for those customers.
- test data sets were all synthetic, launch to real customers fails because data set was not representative of real.
Online:
- yeah we used LLM as a judge in prod/online eval, but the judge sampling failed to measured against our japanese customers, our bilingual customers, customers with media devices etc.
- the LLM as a judge was used, but there was no measure of human accept rate measures/alarms for various use cases (lightbulbs turning off, music playing, etc) so the lightbulbs started failing.
- there was no consistent human sampling of online data, so for example samsung TVs stopped working, but no one caught it until 3 months later when some VP was watching his TV and it broke and was accelerated despite it affecting 2-3% of customers.
I hope you agree people should use the "right tool for the job" meaning a calculator for arithmetic rather than an LLM for everything.
Right job sure. But, If you're replacing the LLM with a jev like as in the article, or switching with a smaller classifier, then you better have good data/evals otherwise you're reducing accuracy without you even knowing.
Also, unless those p90 customers are bringing a big portion of your income (again, if that happens you'd KNOW that your product is broken for them) I'd not pay for the most expensive LLM to "avoid doing evals".
the feedback is part of your online evals... if you have bad feedback you have bad evals
> Also, unless those p90 customers are bringing a big portion of your income (again, if that happens you'd KNOW that your product is broken for them) I'd not pay for the most expensive LLM to "avoid doing evals".
they are but that's besides the point.
Two ways:
- logically ROI of accuracy degradation as compared to cost has to be measured, sure.
- emotionally, morally sub 10% groups in our customer distribution include such things as disabled people, minorities, and the elderly. In my view you should love all your customers.
A harnessed LLM can look this up using tools.
Maybe LLM at least ti prepare context for decision model.
I guess if using a decision model context prep is key.
And what is a decision model exactly? An LLM with instructions to limit responses?
Huh.
But using protecting PII as an example use case is a bit bonkers?
Jev can't answer this for example
You might think it is a contrived example but it is illustrative. It shows that almost most questions are reasoning shaped.But that's only true if we believe the value of 0.85 in itself is correct!? For all we know it could be 0.75 or 0.9 no?
If were to make calls to something like JEV and know that the probability the answer is correct is, say, 0.85193847098713, wouldn't that be even more amazing than JEV and LLMs themselves!?
I'm not being facetious here: how do we know it's 0.85 and neither 0.75 nor 0.9?
And if 0.85 is "correct", why stop there? Why not 0.85193847098713?
Any explanation is very welcome.
Gotcha