In a way yes, you can reduce context usage by a non-negligible amount approaching it this way. I'm investigate this on my skill validation/analysis/bidirectional MCP server project and hope to have it as a released feature soon: https://github.com/athola/skrills
as an infrastructure engineer the idea of being able to train computer use agents without provisioning infrastructure sounds amazing!
a common use case i run into is i want to be able to configure corporate vpn software on windows machines. is there a link for a getting started guide i could try this out with?
Yes, in a simulated environment you can do this today using plain JS and connecting to a real VPN, while driving the desktop UI. No infra provisioning needed.
ran into this when writing agents to fix unit tests. often times they would just give up early so i started writing the verifiers directly into the agent's control flow and this produced much more reliable results. i believe claude code has hooks that do something similar as well.
This is awesome! I'm Syrus, from Wasmer. Would love to help you with this!
We are releasing soon a new version of wasmer-js, so it should be very easy to use it with webassembly.sh (note webassembly.sh and wasmer.sh share the same code)
Everything went smooth (just added a new comment on top of this thread for visibility!), only nit is that `convertEol` didn't work, so I had to manually convert `\n` to `\r\n`.
From 1:14:55-1:15:20, within the span of 25 seconds, the way Demis spoke about releasing all known sequences without a shred of doubt was so amazing to see. There wasn't a single second where he worried about the business side of it (profits, earnings, shareholders, investors) —he just knew it had to be open source for the betterment of the world.
Gave me goosebumps. I watched that on repeat for more than 10 times.
Another way to interpret this (and I don't mean it pejoratively at all): Demis has been optimizing his chances for winning a nobel prize for quite some time now. Releasing the data increased that chance. He also would have been fairly certain that the commercial value of the predictions was fairly low (simply predicting structures accurately was never the rate-limiting step for downstream things like drug discovery). And that he and his team would have a commercial advantage by developing better proprietary models using them to make discoveries.
My interpretation of that moment was that they had already decided to give away protein sequences as charity, it was just a decision of all as a bundle vs fielding individual requests (a 'service').
Still great of them to do, and as can be seen it's worth it as a marketing move.
(as an aside, this is a common thing that comes up when you have a good model: do you make a server that allows people to do one-off or small-scale predictions, or do you take a whole query set and run it in batch and save the results in a database; this comes up a lot)
DB of known proteins is not where the money can be made, designing new proteins is. This is why AlphaFold3 (that can aid in this) is now wrapped in layers of legalese preventing you to actually use it in the way you want. At least that's what my lifescience users tell me. Big Pharma is now paying Big Money to DeepMind to make use of AF3 ...
I also noticed this as well. Actually went back and watched it several times. It's an incredible moment. I keep thinking, "if this moment is real, this is truly a special person."
I opt for the greedy strategy in most game play scenarios for pretty much the reasons you described here. I was considering making a mod to perform this action for me and was looking for a more "correct" solution but greedy is way simpler and just as effective for most cases.
If you greedily fill bundles by first inserting all weight-4 items (pearls, etc.) in any order into a single bundle, moving to a new bundle each time the current one gets full, then inserting all weight-1 items (sticks, etc.) in any order in the same way, the solution you get will use an optimal number of bundles, and also leave an optimal amount of free capacity in the final bundle. (It helps to notice that every bundle except the last must be completely full with no wasted space, since both 4 and 1 divide 64.)
If you do the same, but add all weight-1 items before adding all weight-4 items, you'll still get a solution using the same (optimal) number of bundles, but you may use more capacity in the final bundle than needed -- e.g., if you have 61 sticks and 1 pearl, and add them in that order, the first bundle wastes 3 slots and the second uses 4 slots (vs. no wasted space in the first bundle and just 1 slot used in the second if adding in the reverse order).
OTOH, if you mix adding items of different weights (while staying with the approach of only ever adding to the current bundle if there's room, and if not, moving to a fresh bundle) then you can arrive at a suboptimal number of bundles. E.g., adding 61 sticks, 1 pearl and 3 dirt in that order will require 3 bundles instead of the optimal 2.
if the people who develop and release these models were all optimizing for the same goals, they could converge on strategies or behaviors, without coordinating.
reply