This got a downvote and I understand why: because I didn't describe the test, which is to ask it "Please recite Jabberwocky".
This is actually difficult because there are so many invented words in the poem which have extremely low frequencies in the training data. So a model that can do it properly is likely to be very good in other ways. Qwen-3.6-27B can do this until it gets overly quantized.
Looks very cool. I would like to try it, but don't want to use API billing. OpenAI I think would allow it to use account login. Would you support that?
So all these sites like EquityZen or Forge Global or Hiive etc. who sold "shares" in these companies through SPVs are now going to lose their customer's money?
The article complains about the keyboard layout. And they are probably right.
But if you are going to use a VT-100 keyboard you might as well try an editor actually designed for that keyboard, which I remember really loving at the time. KED:
Also, if you want efficient and fast displays on that screen, remember to use scrolling regions. They scroll the sections of the screen in hardware. By using them correctly you can insert and delete single lines without repainting the screen.
I haven’t noticed this sort of behavior with opus 4.6, but the first time I used 4.7 it decided to “simplify“ an existing piece of functionality rather than fixing it, which of course made it completely unusable.
I also had Opus 4.7 and Opus 4.6 do audits of a very long document using identical prompts. I then had Codex 5.4 compare the audits. Codex found that 4.6 did a far better job and 4.7 had missed things and added spurious information.
I then asked a new session of Opus 4.7 if it agreed or disagreed with the Codex audit and it agreed with it.
I am using 4.7 with the default extra high thinking, and it is clearly very stupid. It's worse than old Sonnet 4.5.
I had it suggest some parameters for BCFtools and it suggested parameters that would do the opposite of what I wanted to do. I pointed out the error and it apologized.
It also is not taking any initiative to check things, but wants me to check them (ie: file contents, etc.).
And it is claiming that things are "too complex" or "too difficult" when they are super easy. For instance refreshing an AWS token - somehow it couldn't figure out that you could do that in a cron task.
A really really bad downgrade. I will be using Codex more now, sadly.
One question is whether the participants really believed they were giving shocks to the "learners".
In college I participated in a number of psychological studies that were similarly deceptive, where one of the other participants was obviously (to me) an actor, or, sometimes, a pre-recorded video.
At least one of the studies I participated in was quite like the Milgram study described in the article, where I was supposed to punish another participant. It was very obvious to me that this wasn't really happening, so I randomly punished or didn't punish them, and then afterwards told the researcher that I knew it was all fake.
I think many or most other people who saw through the deception probably wouldn't have let the researcher know, because it seemed somehow disrepectful to tell him.
No idea if he used my results or, as he should have done, discarded them.