phebbar

curator

[phebbar @phebbar]

We are fast becoming a full ecosystem https://t.co/eGOKyUCf66. Started as a PM and CRM today can run the whole org space.

[phebbar @phebbar]

@PRAVESHPARAS @DILIPtheCHERIAN @sky_phd Well said. I think systemic disciplines like MBA is more relevant now than any time. And one more reason why students need to pursue higher education and just MBA is that they get a different perspective and world view which is essential to a great career.

[phebbar @phebbar]

Wow. This was expected but I did not expect from mid journey.

[François Chollet @fchollet]

Midjourney just announced an extremely innovative new hardware project: full-body internal 3D scans without MRI. Check it out!

[phebbar @phebbar]

@sachdh Well said.

[phebbar @phebbar]

I am feeling more and more that we just have to get back to the days when we built things on our own. Remember Bhabha, Sarabhai, Sir CV Rama et. al. And not wait for grants, tech transfers and the likes.

[phebbar @phebbar]

@ojasvi_yadav May a million models come up now from the land who gave the world the software factory.

[phebbar @phebbar]

How naive of us to think that the powers that be will allow super intelligence to be used by everybody equally without restrictions.

[Sridhar Vembu @svembu]

Our traditional approach to technical complexity is to "divide and conquer" - break the problem into small, low complex chunks and solve them and then put the pieces together. This works well for many problems.

There exists a class of problems that are "fractal-like" in complexity and you break the problem down, only to expose another hidden layer of complexity. Highly creative domains have this nature. That is why they are unpredictable.

[phebbar @phebbar]

@svembu Agreed. We too are working on a way out.

[phebbar @phebbar]

Will try Go today. I have been wanting to code and my muscle memory is gone.

[flowstate @k_flowstate]

banger articles you should read

[phebbar @phebbar]

To build or not to build, that is the question. There’s a skill for every SaaS and there’s an agent for every skill. There’s an SML for every workflow and there is a fine tuned model for every knowledge base.

[Milk Road AI @MilkRoadAI]

Ilya Sutskever just demolished one of the most common dismissals in AI with a single thought experiment.

The standard critique goes like this, LLMs are just predicting the next word and they don't understand anything and are statistical parrots.

It sounds devastating because it's technically true, that's exactly the training objective but Ilya's argument is that people who say this are confusing a task with what must be accomplished to do that task well.

The detective novel makes it visceral.

Picture a novel, hundreds of pages, a complex web of characters, conflicting accounts, scattered clues, deliberate misdirections.

At the very last page, the detective gathers everyone in the room and says, "I now know who committed the crime, and that person is..."

To predict that next word, the killer's name, you don't get to guess.

You can't pattern match on character frequency, you have to have actually tracked every clue across the entire novel, understood every character's motive, recognized which details were planted and which were red herrings, followed the detective's chain of reasoning to its logical conclusion, and arrived at the only name that is consistent with everything that came before.

In other words, to predict that one word accurately, you need something that functions exactly like understanding.

This is the core of Ilya's thesis and it's sharper than it sounds.

The critics are right that the training signal is just predict the next token but if you look at what a model must build internally to do that well across arbitrarily complex text, the required internal representation isn't a lookup table but rather a world model.

It's something that has to encode causality, intent, consequence, contradiction and inference because those are the things that determine what word comes next when you're reading a detective novel at the final reveal.

Research published in early 2025 started formalizing this intuition.

A paper from MIT and other labs demonstrated that next token prediction implicitly recovers human interpretable concepts as latent variables, meaning models trained purely on prediction are spontaneously learning the underlying causal structure of the world rather than just surface patterns.

A separate body of work showed that language models are already measurably superhuman at next token prediction on most text domains, beating human baselines on every benchmark tested.

The debate has moved on from whether next-token prediction can produce understanding.

The frontier question now is whether scaling that process further and augmenting it with reinforcement learning, tool use, and new training methods crosses the threshold into what Ilya calls safe superintelligence.

He left OpenAI in 2024 to found SSI, based on his conviction that this transition is close enough to require a dedicated, safety first organization to get it right.

Predicting the next word is the mechanism but understanding is what has to emerge for the mechanism to work

[Sridhar Vembu @svembu]

Here is what I tell our software engineers on how to thrive in the AI era: be very good domain experts. Programming skills are the foundation (and we definitely don't want to lose them) but deep domain knowledge is what customers pay for, along with reliability, security, support and compliance.

The productivity gains from AI are still hotly debated: we definitely get to a working prototype much faster but a finished product has a lot more to it and not all the stages can be sped up by AI.

That is why I advise our technical teams to not obsess about programmer productivity as a metric but focus on how we can offer a far better experience to customers using AI.

There is a lot of needless or incidental complexity in software that can be eliminated by AI.

[phebbar @phebbar]

Sometimes you got to trust.

[phebbar @phebbar]

Day 2 of #phydi connecting physical real to digital. Actually this goes beyond the ordinary.

https://t.co/oTqyMEJBfM

Drop your comments. And wait for our next leap. #letsbuildtogether
@prakdadlani I had promised this and I will keep doing this.

[phebbar @phebbar]

Code is cheap. The harness and the hardware isn’t.

[phebbar @phebbar]

I kept putting off building a simple postfix based emailing system. Why? Everywhere I read it’s hard to maintain. But I do it anyway. It’s working fine.

[phebbar @phebbar]

This is the first step we took to connect the digital realm to physical. With AI notetaker which automatically assigns work to Bru AI agent which does it right away. @prakdadlani @JayaGup10 @anandmahindra

[phebbar @phebbar]

AI agents are in the digital realm, what if we free them up in the physical world. How should the harness for this be. That is the question we are now grappling with. @prakdadlani @JayaGup10 watch this space https://t.co/vnZIcBKd19

[phebbar @phebbar]

@prakdadlani I like the way you are building things in public. I write software tools now mostly aided by AI assistants. I too will start writing about key stuff and build in public view.