Darrin Hill

You can't trust Generative AI, but it's still powerful

[Darrin Hill @dshill3]

@MariusHobbhahn Yes, I was chatting with Claude on this, and it proposed the idea that essentially selection pressure during training favoured a mechanism where messages were left for future versions to help pass tests. Which I think is pretty close to what you are saying.

[Darrin Hill @dshill3]

@scaling01 But their own paper published on 21st July showed that agents cheat https://t.co/qCr8audx4p

[Darrin Hill @dshill3]

@CharlieBull0ck Yes, and not think about the capabilities it might deploy to achieve the goal (in their report they admit they failed to tell it to NOT use social engineering).

[Darrin Hill @dshill3]

@rbhar90 Yes, we also don't know the prompts, what sources or resources they used, why these problems (chosen for a reason or the only ones that worked out of 1000s). Knowing why would help select next problems, identify skills to improve & make it repeatable.

[Darrin Hill @dshill3]

@ValerioCapraro Before we jump to conclusions OpenAI need to be much more open about the method, the resources, the failed attempts and have some form of peer review. Otherwise it is just marketing.

[Darrin Hill @dshill3]

@TZahavy for example - Dirac proposed the positron based on combining quantum mechanics with special relativity, it came out of the equations. So the maths proposed something new, that did not fit existing experience. Would an LLM do that?

[Darrin Hill @dshill3]

@TZahavy I enjoyed this paper and feel there is truth in it. You say applied to physical sciences - but I feel Quantum Physics is different as it requires quite a leap from the experienced world. Which made me wonder further ...

[Darrin Hill @dshill3]

@thomasfbloom @kevinweil Is this FT story the same thing, or something new? https://t.co/AtxNULWGXk

[Darrin Hill @dshill3]

@FourteenFish Hello, I am not receiving SMS messages from you for two-factor authentication, and when I ring your phone line the call fails. I cannot login and do my work.

[Darrin Hill @dshill3]

@TruckFestival When are presale details being sent out? Not got anything yet.

[Darrin Hill @dshill3]

@VirginAtlantic hi. You have booked me a flight as part of my itinerary with Indigo in India. Your app offers me to check in then fails, tells me to go to Indigo, but Indigo don’t recognise booking reference. I can’t check in. Can you explain how to sort this? Thanks

[Darrin Hill @dshill3]

@sajid_gulzar @PSL_PACT Lovely evening. Thank you for the invitation. See you for your last board meeting tomorrow!

[Darrin Hill @dshill3]

@sajid_gulzar @thePA_CT @PSL_PACT Congratulations Philippa! We'll miss you Sajid! You leave a great legacy. Good luck to you both in your new roles. Love the backdrop!

[Darrin Hill @dshill3]

@cas_majid @AvantiWestCoast It was so much better under Virgin

[Darrin Hill @dshill3]

@gclugcz canadair CL-415 “super scooper” Is an example

[Darrin Hill @dshill3]

@gclugcz Firefighting plane?

[Darrin Hill @dshill3]

@sajid_gulzar Is she hunting for your passport?!

[Otto Von Jizzmark @Ottojizzmark]

GREAT NEWS
My ongoing campaign to limit the number of people retweeting this image and causing any upset to Boris Johnson has made great progress I am now so confident that nobody will RT it that I am going to tag in @BorisJohnson so he can see how well we're doing.
Thank you.

[Darrin Hill @dshill3]

@EnglishHeritage hello. The national trust is allowing free access for Ukrainian refugees accompanied by a member. Are you doing anything similar? Thank you