Avik Dey

Every author writing like this should be required to rewrite abstracts in plain English and read it aloud to an audience of their peers, before they can publish it.

Summary: Conjectural with nice diagrams but no quantitative measures and ignores prior literature.

arxiv.org/pdf/2510.26745

Tom just wants Democrats to be the “responsible adults” while the other party throws childish tantrums. Every time the right drifts further off the edge, he calls for moving the center to meet them - so moderates like him can still feel centered in an unhinged party.

Unfortunately, at this point, any ‘AxI‘ naming is tainted. Whether we use “General”, “Super”, “Hyper” or [Insert], it’s an academic distinction without real world difference. That it also attempts to name a class of models that don’t actually exist, reinforces the conjuring of hype over substance.

LLMs have no utility - is not something I can subscribe to.

LLMs have low utility relative to investment - is my stance.

Utility is tied to cost. Cut LLM spend by 1000x with SSLMs and value equation shifts. Smaller cheaper task-tuned models FTW.

Yes, you still have to pay the humans. Sorry?

Non-monotonic events can still produce a monotonic trend, that’s the outcome we are seeing.

Caring about 10s of thousands of American lives lost in a war isn’t a radical left stance, it’s compassion and empathy.

If you think it’s a far left move, then the Overton window is already badly skewed.

Ahh, you mean the other party moved far right first, forcing the Democrats to move lefter to counter their BS narratives?

Great example of Newton’s Third Law applied to societal dynamics.

Or spitballing here - it just feels like the Democrats moved left, because the other party moved far right?

Karpathy’s tweet is a live demo of the learning loop he promotes. Consciously or not, he is channeling:

- Kolb: Experimental learning theory
- Feynman: Explain in your own words
- Dweck: Growth mindset scale

The medium is the message.

In the research for Computing, my multi-part documentary that examines the intersection of computing and what it means to be human, I've collected almost 6,000 books to help inform my storytelling. You can browse my entire collection here
t.co/fw6RXUYR2l

Don’t have an exact number, but 150+ trick or treaters tonight. One of them:

K (Kid): Trick or treat?
M (Me): Trick.
K: Huh?
M: What’s the trick?
K: You give me candy.

She was the youngest one of the evening. Cutie pie at her best!

Finally …

www.wsj.com/tech/ai/larg...

Yeah, we are saying the same thing - see my previous comment to an earlier post.

I get where you are coming from. Yes we will eventually get to a point where 10x improvement in model efficacy will enable a proportional decrease in dataset size, almost a prerequisite for any semblance of human like intelligence. But that won’t happen till the commercial disillusionment with LLMs.

Sorry, if I am misunderstanding your comment - but, in the field all useful models are intelligent models because commerce says so. Yes, in the lab that’s an important distinction to maintain. In the field, I have never heard of NNs being referred to as anything other than ML, outside of VC offices.

I hear what you are saying and why. However much of the business utility from ML in the 2010s came from scale. Average models trained on massive data often outperformed well designed models trained on limited data. Yes future learning techniques might shift that balance but likely not decouple it.

Yes. I say that because interpreting exactly why the model scored it the way it did is much harder given the highly distributed representations. In fact it’s near impossible task for humans assuming real world internet scale datasets.

Yes - on reliability, but back off on interpretability - even for good old CNN models it’s hard to interpret the how, but easy to verify repeatable reliability.

In their own words:

“Several caveats should be noted: The abilities we observe are highly unreliable; failures of introspection remain the norm.”

transformer-circuits.pub/2025/introsp...

She’s making the classic layman’s mistake of thinking DeepMind is synonymous with AI. If she had actually read even the first paragraph of their paper, she might have clued in it’s a great example of purely statistical machine learning, but that’s probably asking too much.

arxiv.org/pdf/2506.10772

You know how they are going to react to this? Layoff another 100k human engineers to build another super data center because they are convinced they are on the brink of breakthrough - just needs a bit more juice.

Newsflash, boys: AI broke thru back in 2023, now you are just chasing the ghost of AI.

Allow me to summarize:

”If you get elected on a policy stick to it without apologizing. That's what elections are for.”

bsky.app/profile/avik...

When did making kids go hungry become a Christian value?

There is substantial overlap in crypto then NFT now AI investors and it all ties back to their infrastructure investment. Sunk cost fallacy motivates them to keep the bubble afloat.

And on the broader point: it’s not just an absence of alternatives, but also the choice not to build multi-cloud routing into infrastructure design.

True, a deeper issue is lack of alternatives to hyperscalers. But there’s also a narrower design question: “why didn’t Signal hedge with multi-cloud redundancy?“ Even if it compromises NRT guarantees temporarily, degraded fallback seems preferable to a total outage for customers.

Here’s the succinct version:

bsky.app/profile/avik...

The cost of skipping abstraction now will be far worse post integration while quality lags, delays, cost overruns, vendor lock in and the big one - data seepage. Yes, that’s happening.

Build portability now - if you are serious about LLM investment. Abstraction isn’t optional - it’s foundational.

Open source xLMs are rapidly closing the gap with foundation models specially for custom tasks. Stop wiring directly to vendor APIs.

Yes, it might delay your product launch by a few months and that maybe a deal breaker for startups. But if you are an enterprise, you should have no excuse.

On Bsky when posting about scientific papers or articles - I have two modes:

1. Has substance? Put on my scholar hat, assess it in its own register and respond with rigor.

2. Is basic? Hat stays off and I keep it casual because it’s not worth the time.

On “AI“ these days, most fall into the 2nd.

Wait till he gets to the touch problem.