For the core functionality and the optimizations we don't really need to know how the emergent capabilities decide on particular answers.
Which is why we could build LLMs before those features ...emerged for us to see, and why we can just code LLMs with the numerical NN algorithms we use, and do now have to go in and change individual weights.
Me: it’s disturbing we don’t know why this pile of numbers we made seems to *think* in a way previously only done by humans. I think it’s important that we understand this better if possible.
You: we don’t really need to know why that happens.
We do understand "thinking" though. That's literally the whole point of Attention is All You Need, the attention mechanism is what separates the transformer architecture from other neural networks. It's well worth a read if you haven't gone over it yet.
Features like chain-of-thought, long-horizon contexts and RoPE/YaRN all extend this thinking capability very transparently. The only remaining thing to study is the data and weights, which probably isn't going to contain some sort of miraculous revelation.
Attention is at about the same level of abstraction as spike trains or action potential, IMO. It’s a mechanism, it doesn’t tell you much at all about how actual concepts get represented. (It merely defines the substrate with which they can be represented.)
The entire field of Mechanistic Interpretability exists because just understanding Attention does not in any way help you to understand why a certain NN responds with a certain hallucination about a certain Chinese boat in this specific context.
> The only remaining thing to study is the data and weights
To me this is like saying “the only thing left to study in the brain is the connectome; probably going to be boring, we understand it already”. It’s almost all of the hard/meaningful stuff! It’s where intelligence and consciousness lives!
The entire field of mechanistic interpretability has gotten nowhere. The formalized, "causal" understanding of latent space conceptualization isn't any better than an LLM cargo cult.
It's wholly possible that you could study one set of weights for decades, and find nothing. There's no guarantee that any patterns outside of human language exist in that data. In this specific context, it's satisfying enough to state that [Chinese] and [boat] were both tokens in the tokenizer, activated by a feedforward pass through weights that favor [boat] after [Chinese]. There's not any guaranteed solution to this. There's not even any guaranteed problem; that hallucination is an expected behavior.
Sports betting is the one place that "prediction markets" make the most sense. Instead of there being a house to win against, the market just skims a fee. The line is set organically by betters, not be the house targeting a profit margin, and the house has no incentive to restrict successful betters because they solely profit on flow, not losses.
You do have to be careful to avoid letting whales manipulate the perceived market. There are ways to con other gamblers with carefully timed bets.
The bigger the market the harder that is, so maybe it doesn't apply at the level of online sports betting. But organized crime could make trouble at horse tracks.
But if it's parimutuel, don't you lost more on A than you can make on B?
Are you tricking other players, or is the idea that you make lots of bets on A, and then make bets on B fast enough that the bookmaker does not have time to correct?
I think they gloss over a major factor also. They mention:
> Distribution centres are built around motorway junctions (J24 of the M1, the Golden Triangle in the East Midlands) because that’s where road access is.
But they skip _why_ is that road access and motor junction there. It's there because the government decided building roads was something that was it's responsibility. I know this article is UK focused, but for the US if the government decided to build rails also, then they could put them in more convenient places. Instead they allow rail companies to decide which monopoly corridors the companies get to control.
The political solution is 30 year term limits for Senate and House (5 terms and 15 terms respectively). The current system lack of term limits incentivizes inaction
Term limits for elective office are fake, nonworking solution to problems caused by a broken electoral system; the solution is to fix the electoral system, not to impose term limits (which solve nothing.)
Term limits make the problem worse; the main fix is to abandon strict single-member-district first-past-the-post for a more proportional system for legislative elections (for Presidential elections the problem is harder, both because there is no good, easy fix for an inherently single-winner election and because almost any meaningful change will require a Constitutional amendment which is quite difficult even if you can nail down what to do.)
For the House, using a multimember ranked ballots system like Single Transferrable Vote in districts capped at a size of 5 members in states with more than one rep would work tolerably well (especially if combined with increasing the total number of seats beyond the currently-legislated fixed 435.) This does all of support more parties, reduce or eliminate [depending on the exact method chosen] spoiler effects for voting first choice for parties that don't win seats reducing the need for tactical voting, reduce incumbent protection without removing voter choice [because parties are encouraged to run more candidates than they are likely to win], and produce a body that better represents the preferences of the electorate.
The Senate is more complicated because of the 1/3 per class rule, but it can be made slightly better (in order from smallest to largest changes), by:
1. Adopting a single-winner ranked choice method instead of first-past-the-post for Senate elections.
2. Increasing the size of the Senate to three seats per state (electing one Senator from each stare in each of the three two-year classes), combined with #1.
3. Increase the size of the Senate to six (2/state/class) or nine (3/state/class), using a ranked ballots multiwinner proportional system like STV for elections. (3/state/class keeps the majority a significant threshold.
Because of the Constitutional manner of apportioning electors, increasing the size of the House makes Presidential election voting power more equal by population while increasing the size of the Senate makes it less; for this reason, if doing the fixes for Congress discussed above, I would favor not increasing the size of the Senate by a greater multiple than thet of the house, so three per state in the Senate would go with at least a 50% increase in the size of the House, 2/state/class would go with at least tripling the House, 3/class/state would go with at least a 4.5× on the size of the House.
All are very-sound ideas. Regrettably, they'd be tough to explain to the voters — and the vested interests would oppose fiercely.
A good start might be to just triple the size of the House to approximately match the Repr.-to-population ratio when the present 435 number was legislated.
My guess is that they are end-to-end encrypted. And because of Facebook's scale that they're able to probabilisticly guess at what's in the encrypted messages (e.g.a message with X hash has Y probability of containing the word "shoes")
Even “only” information funnels have value if they seek out valuable info, filter, curate. In reality some funnels in this context mutate the message they’re supposed to pass on :-)
Unfortunately it turns the iPhone into a lever that is always trying to launch itself from your hand. The iPhone part is much heavier than the keyboard part. And the ergonomics of the camera control become impossible (unless you have enormous salad fingers or something).
reply