Ashish Bansal

Machine Learning makes predictions. Artificial Intelligence takes decisions. © CEO @StarSparkAI, ex-Google/Twitch/Twitter. / banaras.eth

[Ashish Bansal @ash_bans]

Shared pre-print of my book on #NLP with @data_nerd. Carla was kind enough, took a look & wrote: "Advanced Natural Language Processing with TensorFlow 2 is a great book and it will help out tremendously for those interested in learning more, hats off!!" https://t.co/RSVpkGJQwH

[Ashish Bansal @ash_bans]

@AnthropicAI 's Opus 4.8 sucks. Continuously lies, does not use tools properly to ground assertions and ignores memory/instructions. Examples:
"You're right — my memory even flags that the xxxx is dated and I should use the .... MCP server"
"You were right to call it out. Querying the DB overturned several specific claims in my first answer." (restates all of points it had made)
....I should have led with them instead of filling gaps with confident guesses." after making 14 wrong assertions in a conversation.

Friends, have no fear. #AGI isnt close

[Ashish Bansal @ash_bans]

Really? Isn’t this a free market? If a competitor can copy a business so easily then does the business have a moat?
There is IP protection available if there is a differentiation.
Otherwise there will be many companies with same ideas in the market. It is the hallmark of the free economy.

[Ashish Bansal @ash_bans]

This overloooks the importance of pedagogy and progression. How many parents can highlight the difference between a grade 5 essay and a grade 6 essay?
Say we are teaching factoring a quadratic equation. We don’t start with the quadratic formula as it has a number of pre-requisites. We start with simple quadratics where co-efficient of the square term is one and has integer solutions. Slowly we increase the complexity. This is taught over a couple of grade levels. Adults remember the last most general method.

[Ashish Bansal @ash_bans]

"I just took the same screenshot you showed me, on my machine, with my code applied. It works. Yours doesn't." @AnthropicAI 's Claude Opus 4.7 xhigh after being unable to fix a simple tailwind CSS issue it had introduced.

@OpenAI 's Codex: "The new theme classes trigger malformed generated CSS because tailwind.config.js has broken custom font-size fallbacks for...."
on GPT5.5 High.
Codex 1 - Claude 0

[Songyou Peng @songyoupeng]

Yay, finally! Introducing Vision Banana🍌 from @GoogleDeepMind, our unified model that outperforms SoTA specialist models on various vision tasks!

By treating 2D/3D vision tasks as image generation, we unlock a new foundation for CV.

Project page: https://t.co/GQgRi6mWwC

(1/5)

[Ashish Bansal @ash_bans]

@sama Love codex more than Claude for code generation and tasks. However the recent limits have caused a reduction in my usage.

[Ashish Bansal @ash_bans]

@deedydas Sounds like an LLM wrote this post not realizing how biological development works 😅

[Ashish Bansal @ash_bans]

@NinaDSchick It is a fantastic marketing stunt for sure.

[Ashish Bansal @ash_bans]

@jumperz @peteskomoroch The point being? Sounds like a better google search running locally.

[Ashish Bansal @ash_bans]

@alliekmiller @peteskomoroch Thanks for sharing. We are building AI tutors for educational outcomes. It has been impossible to connect to learning groups at @OpenAI or @AnthropicAI. We get no response at all.
So I have had a bad experience trying to reach out and discuss.

[Ashish Bansal @ash_bans]

@TechLayoffLover Happy to consider them for an ai engineer role @StarsparkAI. They can dm me

[Ashish Bansal @ash_bans]

@whois_legend Please apply using the wellfound link. We review every application

[Ashish Bansal @ash_bans]

https://t.co/RXt7w9xcPc is hiring a UX Designer (Product Design) in Pleasanton https://t.co/qmya0Hz893
We are building AI teachers to help every student learn better.
No recruiters pls, only US based applicants. We are remote friendly. #JobSearch #jobs #AI #EdTech

[Ashish Bansal @ash_bans]

@icelandcricket @ICC You are certainly the T20 champions - Tweet20 that is!

[Ashish Bansal @ash_bans]

@erikalee We can certainly help with the math part at @StarsparkAI , minus the crying and breaking pencils of course

[Ashish Bansal @ash_bans]

@dunkhippo33 At @StarsparkAI we leverage a variety of techniques to make very high gross margins in an #AI native app. But consumer business need funds to grow. So we face a catch 22 when speaking to investors even though we have amazing unit economics.

[StarSpark.ai @StarsparkAI]

Math education is evolving quickly, and parents are doing their best to keep up with a shifting landscape and the growing influence of AI in learning.

To support parents, we partnered with @OpenStax to host a free webinar led by Meg Knapik, Educational Advisor and former Superintendent of Curriculum and Instruction in Illinois, and Ashish Bansal, CEO and Co-Founder of https://t.co/j37n7j4Qbc. They will break down how kids learn math best in today’s education environment and how families can support that learning at home.

In this 45-minute session, we will cover:
• The difference between school curriculum and State Standards
• How Singapore and U.S. math approaches shape student success
• How AI tutoring tools like StarSpark build mastery, confidence, and consistent practice

📅 December 4 at 4 PM CST / 5 PM EST
🔗 Register at the link in our bio

#parentwebinar #matheducation #aiinlearning #mathhelp #mathematics #education

[Ashish Bansal @ash_bans]

@hthieblot Teaches math to kids through AI tutors @StarsparkAI

[Robert Youssef @rryssf]

RIP fine-tuning ☠️

This new Stanford paper just killed it.

It’s called 'Agentic Context Engineering (ACE)' and it proves you can make models smarter without touching a single weight.

Instead of retraining, ACE evolves the context itself.

The model writes, reflects, and edits its own prompt over and over until it becomes a self-improving system.

Think of it like the model keeping a growing notebook of what works.
Each failure becomes a strategy. Each success becomes a rule.

The results are absurd:

+10.6% better than GPT-4–powered agents on AppWorld.
+8.6% on finance reasoning.
86.9% lower cost and latency.
No labels. Just feedback.

Everyone’s been obsessed with “short, clean” prompts.

ACE flips that. It builds long, detailed evolving playbooks that never forget. And it works because LLMs don’t want simplicity, they want *context density.

If this scales, the next generation of AI won’t be “fine-tuned.”
It’ll be self-tuned.

We’re entering the era of living prompts.