Rohit Ghosh

about chasing 'truth' : in quest to search truth, folks are often on the lookout for the objective truth - the one in binaries of right / wrong (founder mode / manager, liberal/conservative, good/bad...) fundamentally, truth is a duality - a cycle in time. right & wrong are popularity contests. you know, it's more like, wave-particle duality. whether it's a wave or particle - is determined by your apparatus for verifying. truth is the duality..

@paraschopra In research at least discovered that novelty and impact are both relative concepts, only relevant in context of other available ideas. Novelty + impact + credibility (of authors in a paper vis-a-vis team in a startup) defines likely arc of the paper/startup.

One of the things the researcher-me is teaching the entrepreneur-me is the importance of doing background search to ensure my project is impactful (in the sense of changing the future), but also overlooked (in the sense of being novel). Novelty + expected impact is the formula.

saw the RMG game ban news in india - thing is you can ban gambling apps, but you can't ban buying hope. hope of an outcome sells better than outcome, especially in a country of billions , if it promises to make you one in a million!

@garrytan not all science labs might need massive dc. if you think about it as generalist + adapter, the depth of adapter might be much thinner for fields like CS or even in medicine (not bio). although over last few month realised the training bottleneck might be curation and not GPUs

In the future academic labs will need massive data centers to run post training and “think hardest” test time reasoning to run custom copilots per lab for original research And if hard and applied science academia doesn’t do it, startup research labs will

want to talk to folks doing active research consumption of AI/ML papers in industry. toying with an open standard document format project. @X , can you please do your job ?

@paraschopra might be out of sync slightly, I think evolutionary algorithms for RL strategy generation / distillation might be key. like mimicking the actual evolution. GAs got a bad repute for long for the compute (for training) - but then compute wasn't so easily available as is today.

When Sam says we will never be able to catch up, he’s essentially right - current paradigm next token prediction on of trillion tokens is senseless to compete on. We must be guerrillas - be on a look out for things that can push us beyond. Head to head completion is senseless.

India should chase the goal of building state of the art AI models that pave the way to AGI, and not simply following the recipe of other labs building LLMs.

in the long flights, why they don't do a few entertainment hours - like maybe in-person performances, sing-alongs etc. ?

@shobhitic @11pointclub @surbhiskjain it was an absolute cracker of an event ! got to meet so many folks :) loved the energy around the place!

from left to right in two and a half months

openAI is down, claude (the free version) is not working....feels like electricity load-sheddings in childhood, internet disruption during college.

being flawless = being mediocre / well-rounded rather tap the crazies, the eccentricities & double down

if renunciation/detachment is for sake of freedom from bondage - is the action even detached from the desire of the result?

all foundational models (open/closed) providers providing AI become like internet service providers providing internet. like services built on internet like e-comm, search - Ai tools would have services/application layers built on foundational models. at the end, the 'legacy' tools would move towards using foundational models as base models (it's quicker for 'legacy' tools to migrate to new base compared to new players gaining distribution. but then there would be exceptions in form of slow legacy players, and fast moving new cos.)

@_rohitghosh Will be interested to hear your thoughts on the impact of open source foundation models will have on “legacy” AI tools.

as building costs go down, distribution becomes the moat which is through 1. product 2. services 3. content the risk of being disrupted : 3 > 2 > 1 the time to gain distribution : 1 > 2 > 3 it's not either/or, most opt for all three. compounding effects play out in all.

@MithileshSaid from @paraschopra x.com/paraschopra/st…

@_rohitghosh This is true, however it takes conscious effort to make "interestingness" the key component. Reality has another dimension There's a vector tugging at the vector of interest that could disrupt the gradient descent - the cost of doing something (time, money, opportunity, etc)

@sriramk @gwern @dwarkesh_sp closest I have found for scientific research papers so far is - openknowledgemaps.org but still not the way I would like it (especially with AI coming in a lot more is possible)

@vistargx absolutely ! drop me a DM

@_rohitghosh Intrested in connecting for a nerding out session on Healthcare ahead?

Healthcare is so broken today. It needs a vision rather than point solutions within sub-specialities to solve it. And I think I got a glimpse of what that vision could like like. Time to clear glasses and see better.

fundamentally, I think it boils down to services you offer. services catering bottom of maslow's hierarchy tends to be paid in money, needs truth seekers like food, security, transport. as you go up the hierarchy, for needs like experience, respect of others - it needs consensus-seekers. it extends to orgs as well outside people.

Truth-seekers take feedback from nature (planes have to fly), free markets (customers have to buy), or competition (militaries have to win). Consensus-seekers take feedback from people (actors want fans, academics want honors, politicians want votes, journalists want status).

if it's a limited pool, there would be bids and hence over time, people with good picks or faculty to tap leveraged skills to produce returns would fundamentally crowd out rest .... but at least the process makes it meritocratic crowding out & kills the 'who got here first' problem for sure !

@_rohitghosh It’s similar but more clear, more direct, value goes to citizens instead of employers, and does way with the “who got here first” system which dominates immigration today.

A market for citizenship: Instead of the government deciding, you, a citizen, pick an applicant from the eligible pool. You get 20% of their tax payments, forever. There’s a secondary market to sell the right to vetting agencies. Pick the next great entrepreneur and retire.