iiinova — where your digital self keeps evolving

Ashish Bansal @ash-bans.bsky.social · 07/10/2026

Is a positive integer between 1 and 3 billion, to be precise

000

Ashish Bansal @ash-bans.bsky.social · 30/09/2026

They are dotting the I in OpenAI

011

Ashish Bansal @ash-bans.bsky.social · 11/09/2026

And the nukes don’t fire themselves. They wait for the button to be pushed by a person.

001

Ashish Bansal @ash-bans.bsky.social · 11/09/2026

I am doing a startup. I get emails to fund my company through family offices. So it may be that there is some personalization going on.

000

Ashish Bansal @ash-bans.bsky.social · 09/05/2026

A clever unit rate problem 😂 However if it is a burrito bowl at chipotle, 40% extra is waste. So I would take 25% off deal in that case. There’s a real world constraint.

010

Ashish Bansal @ash-bans.bsky.social · 07/05/2026

We have to decide if this is a new problem or an existing one. People seem to switch between these two narratives often. If this is an existing problem then we can’t say AI has caused. If AI has caused it then why is there an 80s book talking about it? I think that this is a human problem. Always.

010

Ashish Bansal @ash-bans.bsky.social · 28/04/2026

the world appears to be a moron?

000

Ashish Bansal @ash-bans.bsky.social · 31/03/2026

Oh man. I shudder to think how bad it must have been for a sycophantic model to come back with salvageable. Jokes aside I feel this is a different between Claude and chatgpt. I get relatively “honest” feedback from chatgpt compared to Claude.

010

Ashish Bansal @ash-bans.bsky.social · 31/03/2026

@google Gemini 3 flash preview got a recent update. The model checkpoint seems to be unstable. We use the api version and also see random string of numbers or it going into loops. I hope an update will fix it soon.

010

Ashish Bansal @ash-bans.bsky.social · 30/03/2026

Sorry it took a while :-) @starsparkai.bsky.social 's specialized math teachers are able to explain and plot.

110

Ashish Bansal @ash-bans.bsky.social · 19/01/2026

How do you deal with knowledge cutoff for newer libraries? Also - have you tried codex? I have been impressed by its capabilities in python. Claude I found is a great orchestrator and I use it a lot of automatic end to end testing.

100

Ashish Bansal @ash-bans.bsky.social · 08/01/2026

Aren’t all sports at some level mental? Working tactics, evaluating strategies, having discipline to train, focus, dedication, perseverance are all mental disciplines.

120

Ashish Bansal @ash-bans.bsky.social · 05/01/2026

Nice! Do you store excess in a battery system as well?

000

Ashish Bansal @ash-bans.bsky.social · 05/01/2026

000

Ashish Bansal @ash-bans.bsky.social · 23/12/2025

But we are happy to report that from students in low income neighborhoods of LA learning math in Spanish to girls in rural Pakistan learning math in Urdu, we are grateful that it is making a difference.

010

Ashish Bansal @ash-bans.bsky.social · 23/12/2025

We are focussing on measurable outcomes for students. We do very little marketing. Most of our customers are word of mouth referrals. If it doesn’t make a difference to the student, marketing won’t help anyways 😅

000

Ashish Bansal @ash-bans.bsky.social · 23/12/2025

There are some deeply entrenched beliefs in education. Irrespective of the data presented, these persist. Developers of educational materials and systems need to pander to these to make a sale. So nothing changes. Our math achievement is the lowest in decades.

110

Ashish Bansal @ash-bans.bsky.social · 12/11/2025

My company @starsparkai.bsky.social works in the area of math education. We have a software to generate drawings and graphs and illustrations to power our AI tutors. Let me see if our software can do some of these.

000

Ashish Bansal @ash-bans.bsky.social · 12/11/2025

This is amazing! Thanks for sharing!

100

Ashish Bansal @ash-bans.bsky.social · 11/11/2025

परसाई जी मेरे मनपसंद साहित्यकार है। आनंद लीजिए।

010

Ashish Bansal @ash-bans.bsky.social · 13/10/2025

First time attending NCTM for @starsparkai.bsky.social and me. Looking forward to learning about challenges teachers face in math classrooms.

010

Ashish Bansal @ash-bans.bsky.social · 13/08/2025

Don’t you need to sign in to see them? The screenshot seems to show that the user is logged out…

000

Ashish Bansal @ash-bans.bsky.social · 09/08/2025

Embedding size can be considered as a measure of information. A larger dataset has more information hence needs a larger embedding size to distill that information. However as data set grows information content doesn’t grow linearly due to repetition etc. so embedding size grows sub linearly.

000

Reposted by Ashish Bansal

Ralph Pantozzi @mathillustrated.bsky.social · 07/07/2025

“This paper suggests that the gender inequalities in children’s maths performance aren’t innate or inevitable,” says psychologist Jillian Lauer at the University of Cambridge, UK #iTeachMath www.nature.com/articles/d41...

nature.com

When do girls fall behind in maths? Gigantic study pinpoints the moment

Analysis of almost three million children captures when ‘mathematical gender gap’ first emerges and could help focus efforts to stop girls from falling behind.

071

Ashish Bansal @ash-bans.bsky.social · 14/06/2025

I wonder what the results might be if you put “friendly and rigorous” and “tough minded and kind” as two case. Basically try permutations of the 4 words?

010

Ashish Bansal @ash-bans.bsky.social · 12/06/2025

When the list of tasks is too high it causes brain to get overwhelmed. The granularity of tasks should be moderated by their horizon. Today’s task should be most detailed with a month away tasks high level. I wrote a book with a full time job. What worked was doing my n pages every day.

100

Ashish Bansal @ash-bans.bsky.social · 11/06/2025

I was about to ask what feature engineering techniques did you use - TF-IDF or something else? Also look into BM25… 😜😅

010

Ashish Bansal @ash-bans.bsky.social · 07/06/2025

Net net: you may be safe if you are outside North America and speak a different language than English.

100

Ashish Bansal @ash-bans.bsky.social · 07/06/2025

My experience with voice cloning is that it is not close to my voice. Then I figured out that it is because I have an Indian accent. The voice cloning models have a training data bias for North American accents. Consequently my voice clone ends up with a Spanish accent.

100

Reposted by Ashish Bansal

Saganism @sagan.bsky.social · 30/04/2025

This is from The Tonight Show with Johnny Carson aired on May 20th, 1977. Carl Sagan says something very important, a strong message that didn't lose any validity since then.

411148364058

Ashish Bansal @ash-bans.bsky.social · 10/04/2025

They are not afraid to make mistakes. They are not bogged down by knowledge that “this wont work”.

010

Ashish Bansal @ash-bans.bsky.social · 03/04/2025

What’s the point? They are both empty once I am done with them 😅

000

Ashish Bansal @ash-bans.bsky.social · 16/02/2025

A simple thing we do for our take home assignments : have an interview specifically to discuss the assignment they turned in. Then ask for some modifications to be made in real time. Use of AI as an assistive technology is fine as long as the person has knowledge.

130

Reposted by Ashish Bansal

Weinan Sun @sunw37.bsky.social · 12/02/2025

1/12 How do animals build an internal map? In our new paper, we tracked thousands of mouse CA1 neurons over days/weeks as they learned a VR navigation task. @nspruston.bsky.social & co-1st author @johanwinn.bsky.social Video: www.youtube.com/watch?v=yw_4... Paper: www.nature.com/articles/s41...

youtube.com

Learning produces an orthogonalized state machine in the hippocampus

YouTube video by Weinan Sun

419066

Ashish Bansal @ash-bans.bsky.social · 11/02/2025

We are observing a pendulum swinging back and currently it may feel like it is in the centre. But we all know that without dampening, the pendulum will continue to swing to the other side. So time will tell.

000

Ashish Bansal @ash-bans.bsky.social · 11/02/2025

Can you please share the e link to the paper?

010

Ashish Bansal @ash-bans.bsky.social · 06/02/2025

I have played with the Google’s version only. I felt that the report was lacking specifics or not pulling buffers of info or data from papers and sites it scanned. It always summarized and made the info high level. Don’t have access to OpenAI version. What did you observe on it?

000

Reposted by Ashish Bansal

Melanie Mitchell @melaniemitchell.bsky.social · 30/01/2025

Very good (technical) explainer answering "How has DeepSeek improved the Transformer architecture?". Aimed at readers already familiar with Transformers. epoch.ai/gradient-upd...

epoch.ai

How has DeepSeek improved the Transformer architecture?

This Gradient Updates issue goes over the major changes that went into DeepSeek’s most recent model.

627763

Ashish Bansal @ash-bans.bsky.social · 31/01/2025

For us, prompts are a submodule imported into the main repo. This allows us to independently evaluate prompts and update them. We push them to main repo once satisfied.

010

Ashish Bansal @ash-bans.bsky.social · 24/01/2025

Most of these will be encountered in daily life. 50% off a pair of $18 jeans, or splitting a trip cost of 1268 amongst 4 people.

000

Ashish Bansal @ash-bans.bsky.social · 19/01/2025

After the dissolution of the cartel, the bulbs continued to be manufactured for 1,000 hr lifespan even though they could be designed for much longer operation.

000

Ashish Bansal @ash-bans.bsky.social · 19/01/2025

For a long time, light bulbs were intentionally designed to last only 1,000 hours. Read more about it: en.m.wikipedia.org/wiki/Phoebus...

en.m.wikipedia.org

Phoebus cartel - Wikipedia

100

Reposted by Ashish Bansal

Jay Alammar @jayalammar.bsky.social · 13/01/2025

SWE-Bench has been one of the most important tasks measuring the progress of agents tackling software engineering in 2024. I caught up with two of its creators, @ofirpress.bsky.social and Carlos E. Jimenez to share their ideas on the state of LLM-backed agents. www.youtube.com/watch?v=bivZ...

youtube.com

SWE-Bench authors reflect on the state of LLM agents at Neurips 2024

YouTube video by Jay Alammar

032

Reposted by Ashish Bansal

Sebastian Raschka (rasbt) @rasbt.bsky.social · 01/01/2025

Happy New Year! To kick off the year, I've finally been able to format and upload the draft of my AI Research Highlights of 2024 article. It covers a variety of topics, from mixture-of-experts models to new LLM scaling laws for precision:

magazine.sebastianraschka.com

Noteworthy AI Research Papers of 2024 (Part One)

Six influential AI papers from January to June

26615

Ashish Bansal @ash-bans.bsky.social · 26/12/2024

So is Anil’s criticism. And I am yet to understand your position in this as you jumped into a reply I had posted to Anil.

000