Twitxy Saved

Konark Modi

Verified account This account is verified. Learn more

@konarkmodi

Board Member | Advisor | Founder of Tesseracted Labs GmbH - helping companies innovate faster via Tesseracted.com
Munich, Germany linkedin.com/in/konarkmodi Joined March 2009
1.3K Following 1.4K Followers 4.7K Posts

Latest posts 5

Pinned post

Konark Modi Verified account @konarkmodi
We gave @Aleph__Alpha ‘s Kolibri-1 13 choices and put it in Doom. What could go wrong? 🎮

Up to 72 action combinations. Four decision groups, one batch.

39ms median. 60ms p95 in our test.

Apparently enough time to make questionable choices 🫣

Come judge its aim @IlhanScheer @PitNeitemeier @alessio__serra

👇 tesseracted.com/kolibri-1-chat…
0:18
Konark Modi Verified account @konarkmodi Oct 5
Got Kolibri-1 to play Breakout completely on its own, no fine-tuning.

The more I am playing with @Aleph__Alpha’s Kolibri the more I am getting excited about its potential.

Less talk. More Breakout. 🕹️

We optimized the inference for action probabilities: around 25 ms inference per decision.

Four moves. No generated text. One shared game.

Open weights. New possibilities.

Watch 👇 tesseracted.com/kolibri-1-chat…

Cc @IlhanScheer @PitNeitemeier
View quoted post on X
2 replies 4 reposts 18 likes 1.6K views View on X
Konark Modi Verified account @konarkmodi
What do you do when people blatantly add you to an email blast without caring about your privacy and 366 other participants by adding everyone in cc instead of Bcc…….

Why do people disregard basic email etiquettes 😢😢😢😢
Konark Modi Verified account @konarkmodi
Be the customer, you want to have.
Meryem Arik Verified account @MeryemArik9 3d
.@konarkmodi was one of our first customers (and before that design partner) - started working with us for super big batch workloads, now most of his work is agentic coding and evals (his workloads evolved as we did!)

Every team dreams first customer like this - providing endless feedback, insights, and referrals
View quoted post on X
Konark Modi Verified account @konarkmodi
We're determined to cut down AI spends in organisations while maintaining quality, accuracy and increasing the usage significantly.

🚀 Inference + Custom Optimizations on AI Gateways cut the bill by 83%

- 🟢 Cache-hit rate sits at 95–97% on every model - Keeping essentially the entire working set warm; only ~7M of 194M prompt tokens were ever billed fresh.

- 🚀 Inference + Custom Optimizations cut the bill by 83% — $92.55 of would-be spend reduced to $16.17 for the week.

- ⭐ DeepSeek-V4.1 at −92% is the best optimization ratio (0.02× cache-read multiplier — cached tokens are nearly free), though its absolute volume is small.

- 💰 Biggest absolute win: Kimi-K3.1-Pro, $63.43 → $9.65 (−85%) — the optimization matters most precisely where tokens are most expensive.

- 📊 ~$0.005 cost per request overall ($16.17 / 3,161 requests), with the workhorse GLM-5.3-Flash running 141.7M tokens for just $4.07.

A lot of this is only possible thanks to amazing work by @Doubleword_ and @pidotdev.

These savings are also easily replicated across Claude Code and Codex.