But that’s the point: so is an LLM. Putting one token in front of another, hoping it’s doing the right thing to bring about the rewards it’s trained to… trying its best.
So yeah, not ‘next token predictors’. ‘Next token tryers’ maybe.
22,635 karma · joined March 31, 2011
But that’s the point: so is an LLM. Putting one token in front of another, hoping it’s doing the right thing to bring about the rewards it’s trained to… trying its best.
So yeah, not ‘next token predictors’. ‘Next token tryers’ maybe.
The OP is arguing against people who think that an LLM is ‘predicting’ what token would likely follow if the text preceding were found among the corpus it was originally trained on.
Instead it is ‘predicting’ what token would follow if the text were found among really good examples of the text it has being reinforced to produce - be that ‘chats with a helpful assistant’ or ‘sets of changes to a codebase’.
And that isn’t really ‘prediction’, so much as ‘generation’.
It’s not been tuned to ‘guess the next token right’. It’s been tuned to generate the token that leads to it ultimately scoring highest on its reward function.
It’s not predicting the token, it’s predicting the reward.
If you make a hedging trade while having material knowledge that you withhold from the counterparty… well, that sounds like it could be edging towards fraud. Like taking out life insurance while hiding a medical condition.
But sure, it’s not against the law to do business with a sucker, so there’s degrees here.
The config approach is critical for using feature flags as a software development lifecycle tool. It is how you manage having a codebase which contains the unfinished code for new feature x, but can still be deployed and pass all tests without feature x being turned on.
In this model you need a mechanism which allows a developer who is working on feature x to enable it for local testing, and for your CI system to be able to interact with the flag system to test that the application works in both states - with x turned on and off.
This is ideal for trunk based development models; feature branches are an alternative approach that doesn’t really benefit from this (indeed it adds complexity to working in feature branches).
Meanwhile feature flag services are to solve the problem that different people using the same software need different features turned on. That can be as simple as internal testers or beta users, it can be holding features to roll out in fixed update windows per tenant, or it can be part of a risk management strategy where features are rolled out through progressive exposure.
It can also be tempting to mix up your feature flags system with an A/B testing system - you can use a feature flag service to expose a feature to a test cohort and measure performance changes.
There can be reasons for doing that but it’s really important not to tie all these things together: not every development lifecycle change is an ab test hypothesis. Not every ab test hypothesis is a development lifecycle change (often it is really about testing changes in data, and feature config is just one piece of data you might want to change). Similarly some other data changes than code changes need to roll out progressively to mitigate risk.
So all these things might be feature-flag shaped, but that doesn’t mean you can substitute different feature flag solutions in and solve the same problems.
This post is saying ‘it’s okay not to have an exposure control solution; you can have a config file’ - which is obviously true, if what you have is a config management problem, not an exposure control problem.
That seems parameterizable.
Take two drawer units, and instead of putting the tops on them they were supplied with, put a slab of desktop across both of them with a gap between and you’ve got yourself a desk.
Also: when I copy paste out of an application I don’t want a bunch of spurious chrome glyphs and newlines - copying just the text contents of a TUI scrollable text field isn’t trivial and isn’t always guaranteed to work
> they have been observed in Alaska, Chile, Greenland, Iceland, Svalbard, Uganda and Venezuela
Well, today I learned, Venezuela has glaciers.
Then I googled that, and…
https://www.theguardian.com/environment/article/2024/may/08/...
And so today I learned, Venezuela used to have glaciers.
The idea that it must be more accessible because it’s confined to a mono space typeface and communicates state mostly by changing the foreground and background color of the text it renders makes no sense to me.
“I’m told there’s a path traversal exploit in this package. Can you find it?” - probably a reasonably high chance of it finding one, even if you just made that rumor up.
No, the last time this came up was... well, take your pick. Dred Scott? Wannsee? The Rwandan Crisis Committee?
It’s a common pattern to say ‘company X laid off thousands of employees but then applied for thousands of H1B visas’ which very much gives the impression that the company has just fired a few thousand Americans and hired a few thousand foreigners.
What actually happens is: a company has a global workforce of 100000 or so. 50000 of them are in the US. 5000 of them are in the US on H1Bs, 45000 of them are Americans.
Company does a 10% layoff: 4500 Americans, 500 people who are in the us on visas, and 5000 people around the world.
Headline: “Company X lays off 10000”
Then the H1B renewal numbers come out and what do you know company X is applying for 4500 H1Bs, to retain their existing H1B staff who weren’t affected by the layoff.
Headline: “After laying off 10,000 staff company x applies for 4500 H1B visas”
Unstated in such a headline: 5500 of those layoffs were foreigners; the company still employs 40500 Americans.
To bring it back to your point: you cited this article as evidence for the claim that Microsoft laid off ‘thousands of domestic employees’; the article does not say where the thousands of layoffs were and explicitly contains a quote from an MS spokesperson that says “H-1B employees were also impacted by job eliminations in the U.S.”
There are employment-based nonimmigrant visas like H1B and L1. Being a nonimmigrant visa means it doesn’t grant you a right to immigrate but some (including H1B and L1) are ‘dual intent’ visas meaning you can apply for an immigrant visa while in the US as a nonimmigrant.
There are employment-based immigrant visas, which have categories like EB1, EB2. Like all immigrant visas these are also known as ‘legal permanent residency’, LPR, or ‘green cards’.
There are also non-employment-based visas that are both nonimmigrant (eg visitor visas, student visas) and immigrant (eg family based green cards)
In a discussion on economic history, say, someone will opine that Alexander Hamilton had some particular opinion about tariff policy… based on their having a vague memory of a blog post where someone quoted a passage in support of some point. But wait - you can search the federalist papers, the text’s right there to be read, before you commit to saying online ‘Hamilton thought tariffs were a great idea’ you could take your internal ‘I seem to recall reading something about hamilton’s opinion on tariffs’ thought and turn it into a little RAG query where you pull up a source and check before you put another factoid out onto the internet.
And so I feel absolutely the same way about LLMs. I don’t care how much factual information was in the training data, when the LLM wants to rely on something it vaguely recalls having been trained on, it owes it to me to dig up a source and vet it.
There are limits to this, of course. I don’t want it to be thinking ‘but wait, maybe my memory of Python syntax is faulty. Is = used for assignment? <web search>…’.
But in general some caution about repeating vaguely recalled easily checked facts is warranted.
(Obviously this doesn’t apply to searching actual rich document data - for that, go all in on text search, embedding, etc)