What will be left for us to work on?

[Aaditeshwar Seth]

https://www.cs.princeton.edu/~arvindn/talks/icml-2026-annotated-slides/

--

Aaditeshwar Seth
Microsoft Chair Professor, Computer Science and Engineering, IIT Delhi

Co-founder, Gram Vaani; Co-founder, CoRE Stack

Research | CoRE Stack | Gram Vaani | ACT4D 

Technology and (Dis)Empowerment: A Call to Technologists

Latest@GV: Facebook | Twitter | LinkedIn

[VAIBHAV GARG]

Hello all of you and thank you sir for sharing this topic of what should be left in work for future generation in AI era. I have gone through the PPT but on my honest opinion I think that
the whole PPT was itself manipulation to generate some good PR for AI and it was delivered very shallow. If I go on I could place a counter deep one-liner statement against all the statement of every slide and show that it was shallow. I will comment on it later.

 

In my own experience I am a Fullstack developer in one startup company and our company itself is adopting AI very fast. 
The positive side of it is even though my team of four developers and one designer being reduced to only one developer that is me.

Still I am delivering and taking care of all the software stack that is there in the company, including frontend, back end, mobile app, web app, website, admin panel, WhatsApp communication, their integration with their CRM, ERP, and even developing AI tools equipping Sales, Marketting, Operations and Business Execution team.
And a lot of credit goes to Claud max. $100 subscription. 

So AI is

- Unblocking capablities

- Manifolding scalablity

- And increasing comfort.

But the question is to whom?

The developers who has left the team.So a girl who has left the team. She gave up upon all her hopes that she could be a senior developer someday after seeing that AI already replacing her efficiently. 

Other developers did find the job somewhere else, and designer is daily freeked out by his future scenarios.

Now as per the ppt. I have deep counters against its highlights below

Statement 1: Everytime a new invention comes makes our work better
Ans:Replacing efficient human to do a doable task is not same as replacing inefficient machine with an efficient one - everyone knows the game here is cut the cost.
Statement 2: Machine takeover of human job is not new and scary.
Ans:Previously invention replaced primary adaptable geeky labour, now it is replacing tertiary domain expert employee.
Statement 3: Writing code was not bottleneck
Ans:No bottleneck was the cost of qualified human it comes with ablity to write the code that scale businesses to make them global.
Statement 4: AI amplifies human potential but the human remains in control.
Ans:Only one human remains in game to whom responsibility of every AI action can be dumped.
Statement 5: It opens new avenues of solutions
Ans:Yes its opening new avenues - all new majorly capitalist avenues extending more control over public money and life.
Statement 6: Lawyers a good example to use AI in judicial expedition
Ans:Yes this is the problem to focus. But will AI do human good - if behind it is not capital good that control and decides the use cases of AI.
Statement 6: Not a super intelligence
Ans:Its super because the number of neuron it uses now is more then 100 times I have in my head and it works non stop 24×7 while I work at max 8 hours.
Statement 7: Human Representation quality is absolutely fundamental to cognition
Ans:Representation is just a presentation of Information know the absorbability of audience who are targets. Soon AI being exposed to social media will perfect that too.

Here it may seem that I am against AI but it is not. I am even in the support of AI but I am just asking for whose motive AI is in work. 

Whose action and whose aim gonna control the AI in next decade will be controlling the humanity itself.

If we think of real problem statement of humanity today, either it's garbage collection and waste disposal, either it's pollution, either it's global warming, either it's the work, the problem for which Co-Stack is providing work, eiither it's about corruption, either it's about democracy, either it's about bringing more transparency to Legislation, to Executive, to Purocrisi, to Judiciary, to Civil Institutes like Hospital, Educational Bodies, Tesis, Blocks, Talukas. 

I know this won't be the focus of the time even with AI because nobody is putting their money to churn out the system which is not producing revenue for some capitalist benefit. There we will be listening the terms like that AI too expensive to be applied to these use cases. It is using a lot of water and electricity and other resources and depriving humanity and all. Whereas in a parallel world AI will be kept using for generation of dumb entertaining videos and content that serves no human benefit rather than to manipulate people in putting their money, time and mental energy into capital gains of someone on the top.

An optimistic AI use case:

Take the FMCG sector. Most products have little real IP—the advantage comes from branding and distribution. Today, AI can help anyone build such a business by handholding him in legal registration, compliance, accounting, supply chains, operations, and management.

Now imagine a policy that limits non-IP FMCG companies to operating within just three districts. AI could enable thousands of local businesses, creating more jobs, competition, and consumer choice.

The problem isn't AI's capability—it's who controls its use. And policies alreay in place to support game of large corporate interests.

That's why I believe AI will mostly serve those who already have capital and influence, thats why world fear that they will create AI into some super Intellegent monster in their selfish race to the top.  

I will like to invite all rectifying or affirmative feedbacks for my perspective on it, warmly.

Regards,

Vaibhav Garg.

[Shubham Kumar]

Position paper: LLMs can't jump | Jan 2026

It argues that LLMs have become good at language generation and other forms of media but of the three forms of cognitive reasoning: induction(generalising from examples, function learning, eg: supervised learning), deduction(deriving specific facts from primitives and general rules, eg: logical reasoning and mathematical proofs), and abduction(making leaps to find better structures and frameworks to describe an existing phenomena; no such ability observed in LLMs), LLMs have become good at induction and now at deduction but abduction still eludes them.

Also, the LLMs' symbolic manipulation is not grounded in physical reality. So, whatever it is generating and performing is more easily explained by the data distribution on which it was trained rather than its generalised cognitive ability(which is yet to be developed). See, the research in 'grounding in ai' which is actively being pursued in Neuro-Symbolic AI community(for example: Prof. Parag Singla, Prof. Rohan Paul, both at IITD have active research work in grounding AI) and research on World Models, which is aimed at multimodal reasoning.

Statement 6: Not a super intelligence
Ans:Its super because the number of neuron it uses now is more then 100 times I have in my head and it works non stop 24×7 while I work at max 8 hours.

Regarding this statement:  Large number of parameters and components means larger complexity(could also mean inefficiency). I agree that we have only ~86 billion neurons and the largest models have around trillion parameters. But, at first the active parameters are low because all these frontier models are Mixture of Experts(MoE), so at a time only nearly a fraction(say one-tenth)  becomes active. Also, there are many systems that have a larger degree of parameters and complexity, for example: power systems(largest man made systems). LLMs are attempts at mechanising intelligence in the same way as computers and calculators were attempts at mechanising computation. Once something is mechanised, it can be made fast -- whatever aspect is finally mechanised, that becomes fast. Does running a program for 60 days or 1 year non-stop worries anyone? No. So, if present LLMs remain in this state and work non-stop then it would definitely do the heavy lifting of many aspects of data visualisation, report or presentation generation, front-end, backend, theorem proving, pipeline design, creating workflows, writing software, etc. 

Unless LLMs become good at abduction(as explained earlier), become good at finding the problems that need to be worked on, and are able to create framework or structures, knowledge workers still stand a chance. 

What I also feel is that the present hardware designs are inefficient and better for example in-place memory and computation(even improving on Single Instruction Multiple Thread processing paradigm of GPUs) or neuromorphic chips could help design better systems. 

But I like many other points you raised. 

Best Regards,

Shubham Kumar

 

--
You received this message because you are subscribed to the Google Groups "core-stack-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to core-stack-de...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/core-stack-dev/07d83407-eece-49d5-907c-ef18995b6488n%40googlegroups.com.