The stop using non-deterministic technologies to do tasks requiring deterministic outputs challenge.

[securitywriter@infosec.exchange]

The stop using non-deterministic technologies to do tasks requiring deterministic outputs challenge.

Let’s try that.

[rachel@transitory.social]

@SecurityWriter@infosec.exchange this is something I have to try and explain in meetings on a weekly basis.

I've heard some terrifying suggestions for usage of LLMs

[mttaggart@infosec.exchange]

@SecurityWriter This is short-sighted imo, and missing a huge monetization opportunity.

What if every function call, every API endpoint, every program output could be gambled on in a betting market?

Will my function return the desired result? Current odds are 4:1!

I just checked and polymorphic.market is available. Let's do this.

[pikesley@mastodon.me.uk]

@mttaggart @SecurityWriter step away from the lathe please

[cr0w@infosec.exchange]

@mttaggart @SecurityWriter The worst part is how many people would call that a brilliant idea.

[theorangetheme@en.osm.town]

@mttaggart @SecurityWriter Now I want to make a website that just spits out the next number from /dev/urandom, while people place bets on what it'll be. I'm pretty sure I could write this in an afternoon.

[theorangetheme@en.osm.town]

@mttaggart @SecurityWriter I've been informed that this is just roulette.

[mttaggart@infosec.exchange]

@theorangetheme @SecurityWriter The roulette is more random

[ckure@infosec.exchange]

@SecurityWriter I get where you’re coming from, but what about trying this new thing? It comes with a free* frogurt!

*Offer valid while surprise lasts

[securitywriter@infosec.exchange]

@mttaggart Process Pachinko.

Every ‘bucket’ has an outcome. The jackpot just gives you more balls.

We could get some sweet animated dashboards, a Dynatrace integration which will show all green anyway, some sort of SIEM runbook which naturally make things worse, glossy exec reports.

Stick some GenAI in there, so you can ask it to influence outcomes and it be utterly incapable of doing so.

Add in premium modules for rate limiting, a rabbits foot, money paw, four leaf clover, ball credits (not getting the successes you need? Try more requests!)

We can spin up a side hustle which allows betting on business outcomes, which definitely isn’t gambling, but ‘investing’.

Honestly, we’ve just won the cloud. I was so wrong.

[securitywriter@infosec.exchange]

@mttaggart I forgot the damn partner program.

For each new customer you refer to us, you get a 0.5% probability boost to your favoured outcomes*

For an additional $325k/yr you can get a spotlight announcement and a seat on our board of directors with voting rights.

*Annual quotas apply to maintain participation. Limited to maximum 120% probability per outcome.

[securitywriter@infosec.exchange]

@mttaggart make this an incremental game and I’d play it.

*Proceeds to stave off overwhelming ADHD impulse to drop all current projects and learn how to make this happen.*

[realn2s@infosec.exchange]

@SecurityWriter @mttaggart
You should also add some non-deterministic input. Promoting can hook users and at the same time you can always blame the user "if the results are incorrect you are prompting it wrong"

[transcaffeine@mond-basis.eu]

@SecurityWriter "i asked chatgpt and it said that it's fine and i should stop worrying about details"

[drahardja@sfba.social]

@SecurityWriter I witnessed someone saying that their “agentic” AI doohickey does what you ask it to do 85% of the time, and their friends celebrated that as an accomplishment.

The bar is so very low.

[monniauxd@social.sciences.re]

@SecurityWriter stuff everything in a may-return monad and show convergence to a deterministic result to exit the monad!

https://hal.science/hal-04859400

[star@fed.amazonawaws.com]

@SecurityWriter As I understood it, LLM output is technically deterministic, but because of the random seed thingy it gives different answers each time, no?

[securitywriter@infosec.exchange]

@star the randomness is guaranteed 🤭

[star@fed.amazonawaws.com]

@SecurityWriter well, the wrongness certainly is, and, FYI, i am against ai; i just thought that this was how it worked: same seed, same input, same context = same output (on same model), or am i mistaken?

[exchgr@mastodon.world]

@SecurityWriter @drahardja difficulty: impossible