If LLMs are not unlocking additional revs and depth, albeit making it all net slower in the short term, are you even working ?
And after working for sometime, you'd definitely know you want more of this, but perhaps in a way that is easier to use, which would be net faster.
If you start with the assumption that people think on a spectrum, it’d be safe to assume that around 50% people would hate you.
Accepting this makes life super easy. It’s much easier to make the judgement who should hate you - cowards, hardcore hypocrites, turncoats etc
I have taken so many investor meetings from the car, given the lack of infra, the heat, the traffic and travel times, and the need to get multiple things done !!
It should not be embarrassing. The poverty & struggle of most people is much more so. Ivory Towers !
Lawyer appears before Supreme Court sitting in her car:
Court: you are a responsible counsel. Does it behove a counsel sitting in the car and appearing?
Lawyer: I have come for a mediation case
Court: you could have gone inside, sat in a room. We feel embarrassed. I don’t know how you don’t feel it. We are feeling embarrassed for you.
Last night on @Ndtv, I asked the NTA director, courtesy @ShivAroor. There are 5 questions that bother me on the NEET paper leak and I asked:
1. Public MBBS seat costs about Rs.5 lacs while a private one costs Rs.1.5 crores. The financial incentive for families and brokers is far too high. How do you counter the imbalance?
2. NEET is pen & paper exam. The physical paper is transported to 5000+ Centers and at least 200,000 hands. The time gap is often weeks before the exam. How do we solve this?
3. Only NTA knows the extent of the leak. The fact that they cancelled the exam means it is substantial. What was the extent of the leak?
4. Was anyone punished for the 2024 leak fiasco? Can we have a white paper on what we learnt from 2024?
5. Can India reduce the dependence on a single entrance exam that determines the fate of students? If we have 50% weightage for school scores, we can curb the coaching mafia menace..
Basically latency impact partly is covered in fluency and efficiency factor, and error rate in the intelligence factor.
I would like focus on this part - efficiency of common encoding language, or probably the efficiency of recursively breaking the context.
The question is how many parallel threads can I handle ?
Depends on the horizon length of the time AI takes to do a task, and probably my ability to context switch.
Probably a mathematical upper bound can be found, that is related to the memory & intelligence of the entities, and the fluency in and efficiency of common encoding language. Probably also on the steps (iterative and recursive) required to achieve the goal.
i continue to not be a fan of the direction model apis are headed. i think the beautiful thing about chat completions is that it does exactly one thing -- messages in and messages out -- and leaves the rest to the creativity of a developer.
the next gen of APIs have lots of specific, built-in flags for specific features (web search, caching, reasoning) and lots of portability issues across providers.
i genuinely do not believe the motivation is lock in (even though it seems suspicious), and instead i think they're just trying to get us access to new stuff that has an evolving surface area. but a swiss army knife api with 100 knobs is hopefully not the final solution.
i'm hopeful that a new, appropriately extensible, pattern emerges. i think the best abstractions are extremely simple and flexible on the surface, with a lot of depth underneath. the current vintage of apis.... do not fulfill that premise.
CLI is closest to what the model actually sees.
Models are surface learners, and the whole development process can be packed and unpacked as a text surface with CLI commands.
CLI is natural interface for the model, and if it is doing 99% of the lifting, answer is obvious.
I can’t believe I have to be the one saying this:
An IDE for software development is better than a command line interface.
We fought this war and won it a long time ago.
Why is now every model provider releasing a CLI for AI coding?
What’s this madness about?
@malpani Hard Disagreement.
The role of a coach is to teach the art of combat, not to choose the battlefield. Expecting tuition to also supply life direction is like blaming a swordsmith for not declaring which enemy you should fight.