The problem with that perspective is that people thought, "Only AGI can do X, therefore, if a thing can do X, it's AGI." Because they can't imagine how X could be accomplished without it.
However, what's actually changed is how people perceived X because we don't have to imagine. We understand now that it doesn't require AGI so we no longer make that leap to assume it's AGI if it can do X.
It's really going to be a "I know it when I see it" situation.
They are instructions. Everything in the context is instructions for the next token. The "thought" guides the answer by providing clearer instructions.
I'm not sure where you're getting accurate data from (I couldn't seem to find any that seemed trustworthy) but Anthropic and OpenAI, being the two big players, presumably take ~40% and 5% isn't that much of a disadvantage in a market where users can move rapidly.
Given the quality of the last couple Grok models, I'm sure their share is on the upswing as users make the switch.
> If the new result is real, more signals should emerge soon. LZ researchers have already collected three times as much data as they used in the paper.
It sounds like this implies they've seen 3x more events but it seems like they would have said that if it were the case. Have they just gathered more data about the single event or is this 4 separate events they're talking about?
Detectors like this work on exposure. They're always on (except for calibration and maintenance), waiting for events to happen. This paper was written with 2.8 tonne-years of data. That is, 4.7 tonnes of liquid xenon for a little mmore than half a year. The detector has 7 tonnes, and the 4.7 number reflects cuts they made on parts of the detector that either they don't understand as well, or have higher backgrounds.
As they better understand the detector, they can use more of that mass. They have data from it, but they just didn't use it. And they're always collecting more data, too, as time passes.
So the 3x is saying they have something like 8.5 tonne-years of data.
Hm, they have 7 tonnes of Xenon. Events detected all around in the matter, but the PMTs can localise where the event happened. So they can virtually segment parts of the detector where they are sure all the outside effects are understood and taken care of.
If this anything like CERN detectors, they get amounts of data so vast that they have to discard almost all of it to be even able to record it. Depending on heurestics you use to discard data you might be discarding what you are looking for and after adjustment will get some new interesting events, but still actually processing the candidates might take a long time.
This raises what is (I think) an interesting question. CERN is a collider, so they are _trying_ to produce lots of stuff, and they do (lots and lots of stuff). They can't write it all to disk, and most of it isn't interesting enough to try.
The work being done here falls into the category of "low background physics" --- they aren't trying to produce anything, and actually put quite a bit of effort into doing the opposite, by removing all sources of particles (e.g. sourcing materials free of radioactive contaminants, physically cleaning all surfaces and purifying all fluids involved, etc).
So the detector, if built properly, is fairly quiet, and you try to write as much data to disk as you can (e.g., if something even fairly-potentially interesting happens, you save it). Then when you analyze the data like this, you ignore the majority of what you've got --- only a teeny fraction makes it into an analysis of this caliber.
They collected x3 more hay, and they still have to processes it and try to find any needle mixed with it.
Hopefully it the new data may have 3 additional events, or perhaps 2 or perhaps 4 or perhaps 10 or perhaps... Or the reported event may be false event caused by a lucky coincidence, and they may find 0 additional events.
Whether or not they've "seen" 3x more events is a little bit of a tricky question, because while they may have captured 3x the data exposure (see sibling comments) experiments often operate blinded to the data. They can develop their analysis scripts, play out various different scenarios via Monte Carlo simulation, and get their whole pipeline working without the bias of actually seeing how each change in algorithm alters the outcome for the real data.
Then, at some point, they freeze their pipeline, "open the box", run the analysis on the real data, and report what they find. But they can only "open the box" once per exposure, after that you can worry that human bias can creep in.
Then it’s not a benchmark. I think his thesis is solid: if neither pelicans nor bicycles stick out, it follows that there isn’t special attention being given to them by the labs
Thank you. This is what I'm driving at, that most of the AI and software devs here seem to be missing. Intelligence is not, and never was the bottleneck for most science/hard tech. Full AGI gets, at best, a small productivity improvement, which over long periods of time does have compounding effects. But this isn't a singularity hard-takeoff inflection point.
I had never heard of scratch before but I find this line of thought generally backwards. Maybe that's what carmack is getting at but it seems to me that early "exposure" doesn't really matter as much as people want it to.
There is so much software all around us these days, any curious kid who's interest is piqued by it will find a way in to the hobby and maybe to a career.
There was a period where having a computer in a classroom was a big deal because it might be a kids only exposure to one but that's not the case today (not everywhere, I know). Kids pursue their interests if they can and software is very easily pursuable these days.
There are many, many mountains to climb. Some are harder than others. Each doesn't need to have an escalator so everyone can climb it.
Developers have limited time and all features, including accessibility features, take away from that time. There are many games I never beat because they were difficult.
