

Interesting point @13:12 of the video though; agents showed good performance improvements in limited scopes, but human researchers still showed much better improvement, over a factor of 2, ‘…showing better returns for increasing time budgets’.
There may be evidence of some ‘recursive self-improvement’ in research models rigged to modify and re-evaluate themselves, but it’s still only in very tight, limited knowledge areas. This isn’t SkyNet.










noai.duckduckgo.com, for at least one alternative.
CoMaps or Openstreetmaps instead of Google Maps.
ProtonMail/ProtonDrive instead of gmail, gdrive.
There are lots of alternatives, we just need to use them and tell others to use them as well.