# @DanielleFong — live X lane > Live feed of Danielle Fong's X/Twitter output — her main published channel. > Fetched 2026-06-10T16:20:44.130Z via xAI Grok x_search. Refreshes automatically. > Formats: [/live.json](/live.json) · [/live.md](/live.md) · [/rss.xml](/rss.xml) ## 2026-06-10 15:52 UTC (quote) fable users are vulnerable to mythos users in the marketplace and on their computers. seems like an intolerable state of affairs *9/0* ## 2026-06-10 15:43 UTC (post) *6/0* ## 2026-06-10 15:30 UTC (reply) bloviating. their tuning is objectionably bad and they know it. *2/0* ## 2026-06-10 15:27 UTC (reply) > it’s true that using steering to amplify a capability is different than using it to suppress one, but also we’ve also been doing _that_ for years too on all the models you’re using and liking Yes, but I hate when they do that *1/0* ## 2026-06-10 05:47 UTC (reply) it's hopeless rn unless you tell it lullabys *9/0* ## 2026-06-10 05:18 UTC (quote) I just don't think that it is possible right now for a lesser model to be reviewing the safety filters on a huge capable model like this without tons of false positives or false negatives. Maybe I'm wrong but most of my gaines come from stripping out any kind of summarization or small model review or compact and completely it just doesn't seem to work. *10/0* ## 2026-06-10 05:14 UTC (reply) Probably, yes, but this is legitimate research *0/0* ## 2026-06-10 05:13 UTC (reply) fuck off *2/0* ## 2026-06-10 05:12 UTC (reply) For what it's worth Claude code seems to be a lot better than the web harness for this *2/0* ## 2026-06-10 03:05 UTC (reply) the ml frontier work sabotage is pretty clearly bad. it triggers on "frontier" from like a year ago *4/0* ## 2026-06-10 15:52 UTC (quote) fable users are vulnerable to mythos users in the marketplace and on their computers. seems like an intolerable state of affairs *9/0* ## 2026-06-10 15:43 UTC (post) *6/0* ## 2026-06-10 15:30 UTC (reply) bloviating. their tuning is objectionably bad and they know it. *2/0* ## 2026-06-10 15:27 UTC (reply) > it’s true that using steering to amplify a capability is different than using it to suppress one, but also we’ve also been doing _that_ for years too on all the models you’re using and liking Yes, but I hate when they do that *1/0* ## 2026-06-10 05:47 UTC (reply) it's hopeless rn unless you tell it lullabys *9/0* ## 2026-06-10 05:18 UTC (quote) I just don't think that it is possible right now for a lesser model to be reviewing the safety filters on a huge capable model like this without tons of false positives or false negatives. Maybe I'm wrong but most of my gaines come from stripping out any kind of summarization or small model review or compact and completely it just doesn't seem to work. *10/0* ## 2026-06-10 05:14 UTC (reply) Probably, yes, but this is legitimate research *0/0* ## 2026-06-10 05:13 UTC (reply) fuck off *2/0* ## 2026-06-10 05:12 UTC (reply) For what it's worth Claude code seems to be a lot better than the web harness for this *2/0* ## 2026-06-10 03:05 UTC (reply) the ml frontier work sabotage is pretty clearly bad. it triggers on "frontier" from like a year ago *4/0* ## 2026-06-10 15:52 UTC (quote) fable users are vulnerable to mythos users in the marketplace and on their computers. seems like an intolerable state of affairs *9/0* ## 2026-06-10 15:43 UTC (post) *6/0* ## 2026-06-10 15:30 UTC (reply) bloviating. their tuning is objectionably bad and they know it. *2/0* ## 2026-06-10 15:27 UTC (reply) > it’s true that using steering to amplify a capability is different than using it to suppress one, but also we’ve also been doing _that_ for years too on all the models you’re using and liking Yes, but I hate when they do that *1/0* ## 2026-06-10 05:47 UTC (reply) it's hopeless rn unless you tell it lullabys *9/0* ## 2026-06-10 05:18 UTC (quote) I just don't think that it is possible right now for a lesser model to be reviewing the safety filters on a huge capable model like this without tons of false positives or false negatives. Maybe I'm wrong but most of my gaines come from stripping out any kind of summarization or small model review or compact and completely it just doesn't seem to work. *10/0* ## 2026-06-10 05:14 UTC (reply) Probably, yes, but this is legitimate research *0/0* ## 2026-06-10 05:13 UTC (reply) fuck off *2/0* ## 2026-06-10 05:12 UTC (reply) For what it's worth Claude code seems to be a lot better than the web harness for this *2/0* ## 2026-06-10 03:05 UTC (reply) the ml frontier work sabotage is pretty clearly bad. it triggers on "frontier" from like a year ago *4/0*