With Zlide.io, you can research any topic, turn your findings into a stunning deck, and export it as a videoโall powered by highly optimised infrastructure. ๐
For reference, generating this video would cost us around $0.50, compared to $5โ$10 on other video engines offering comparable capabilities.
Frontier labs shipped again this week โ and the pattern is clear.
Small models got dramatically cheaper. Sparse MoEs keep scaling total parameters while keeping active compute low. Open weights and gated cyber models are pulling in opposite directions #AI
RSI is powerful but overfits. 93.0 on evolve can be 40.6 OOD.
RRSI regularizes HOW you search, not WHAT you edit: annealed sparsity ๐ท,
leakage critic,
noise floor ๐ท,
cost-aware ๐ท, +
pruning.
+14.1 in-dist, +4.7 OOD, 30% fewer tokens.
Paper: alphaxiv.org/abs/2609.24972
Video ๐ Zlide.io @zijus
In 2022, ๐๐๐๐ฌ ๐๐ญ๐ ๐๐๐.
Remember the stack? ๐๐ฆ๐ฏ๐ต๐ช๐ฎ๐ฆ๐ฏ๐ต, ๐๐๐, ๐ด๐ถ๐ฎ๐ฎ๐ข๐ณ๐ช๐ป๐ฆ๐ณ, ๐ช๐ฏ๐ต๐ฆ๐ฏ๐ต ๐ค๐ญ๐ข๐ด๐ด๐ช๐ง๐ช๐ฆ๐ณ - 4 trainings, 4 pipelines.
Then one prompt replaced all four: `Classify sentiment, extract entities, summarize`.
90% of custom NLP pipelines disappeared. Not because BERT was bad, but because semantics became API-callable.
I think we're seeing the exact same unbundling with classification.
๐ณ๐ณ๐ด : ๐ต๐ณ๐ท :: ๐ฑ๐๐ : ๐ช๐๐๐๐๐๐๐๐๐๐๐๐๐
Jev isn't a chatbot. It's a System One model - a massive, calibrated classifier: ๐ช๐๐๐๐๐ as probability, not text.
๐๐ฉ๐ฐ๐ช๐ค๐ฆ. ๐๐ค๐ฐ๐ณ๐ฆ. ๐๐ฐ๐ถ๐ญ. One call, 200ms, parallel. Code owns the workflow; Jev supplies the prediction.
So the question for everyone who still trains LightGBM / XGBoost for lead scoring, ticket tagging, routing, guards:
2022: If you can write the label in English, you don't need to train an NLP model.
2026: If you can write the label in English, do you still need to train a classifier?
What's your take - what dies and what lives?
#MachineLearning #LLM #Jev
@typesafeai
Slide Deck zijus.com/zlide/dhs
What an incredible start to #DataHackSummit2026! ๐
Our Co-Founder, Sonny Laskar, took the stage with a live coding session:
๐๏ธ Killing the Awkward Silence: Architecting an Ultra-Low-Latency Voice AI
Stop building AI agents with duct tape. ๐งต
We are launching @zijus : The worldโs first Full-Stack PromptCode platform.
๐๏ธ Infrastructure
๐ง Context Graphs
๐จ Generative UI
๐ Omnichannel (Voice/WA)
Everything is a prompt. ๐ช
zijus.com/try #agents
โจ Open-sourced DiffXcel - visually compare Excel files in your browser!
Color-coded diffs โข Sheet navigation โข Exportable reports
๐ Python/Flask โข ๐ github.com/sonnylaskar/DiffXโฆ
#Excel #OpenSource
Honoured to receive the ๐๐๐๐๐ ๐จ๐๐๐๐๐ ๐จ๐๐๐๐๐๐๐ ๐จ๐๐๐๐
- 2024 (๐น๐๐๐๐๐ ๐บ๐๐๐) from @IIM_I ! Grateful to my mentors at @MicrolandLtd for their guidance #IIMIndore #Gratitude #Leadership #Growth
Today the largest known prime number was discovered. We will have to wait for few more years to find the next. Mathematics is amazing ๐
A recent @capitalmind_in report highlights that most #MutualFund investments are still funneled through distributors. This suggests a significant gap in understanding, as many investors could benefit from the lower costs of direct funds. Time to educate ourselves and explore the direct route! Thanks @deepakshenoy for this wonderful piece of material.
Celebrating a decade at @MicrolandLtd, as the company marks its 35th anniversary! It's been an honor to contribute to AI product development and engineering during such a significant milestone. Grateful for the opportunities, the team, and the memories. #Microland #AI #ProductDevelopment #Engineering #anniversary
Ratan Tata, a true titan of industry, has passed away at 86. His legacy includes transforming the Tata Group and many more. Rest in peace. #RIPRatanTata
Don't put SL orders in the system, such spikes will hunt your SL orders and come back #stockmarkets #intradaytrading #finnifty














