Lossfunk Letters

Your LLM won't believe you, but it WILL believe anyone "verified"

Models that shrug off a wrong user still flip for the same claim from a "verified source." We call this Authority Bias.

Oct 1 • Abhinav Rajeev Kumar

9

2

1

September 2026

Traveling into the future

Reflections on 2 years of Lossfunk and the path ahead

Sep 7 • Paras Chopra

33

12

3

August 2026

Diffusion models can see an illusion. Then they forget it before the picture comes out.

How denoising models quietly build human-like illusion representations deep inside the network, and throw the signal away before it ever reaches a…

Aug 3 • Lossfunk and Gautam Ranka

3

2

June 2026

LLMs are blind. But somehow they can see.

How text-only models develop human-like maps of color, pitch, emotion, and taste, without ever experiencing any of them.

Jun 30 • Lossfunk

8

1

1

The Metaprogramming Reflex: How the Best Coding Agents Survive Languages They've Never Seen

The agentic follow-up to EsoLang-Bench, now with tools, a workspace, and one strategy that separates the strong agents from the rest.

Jun 18 • Aman Sharma and Lossfunk

4

2

May 2026

How to make progress on metaphysical puzzles in AI?

New paper from Lossfunk

May 21 • Paras Chopra

5

3

1

Beyond Reward Design: Discovering RL Interfaces with LLMs

Jointly evolving RL observations and rewards with evolutionary LLM-guided search

May 11 • Akshat Singh Jaswal

5

3

April 2026

Attributes of a great research question

Sharing our learnings from iterating on what good science is

Apr 29 • Paras Chopra

24

2

10

Can AI models be conscious?

How can we tell?

Apr 21 • Paras Chopra

12

1

3

Does spatial context make VLMs better game-playing agents?

And why noisy perception can make them worse.

Apr 2 • Ashish Baghel

1

1

2

March 2026

The Reasoning Illusion: Why LLMs Fail When the Training Data Runs Out

EsoLang-Bench — accepted to the Logical Reasoning and ICBINB workshops at ICLR 2026

Mar 19 • Aman Sharma

9

1

2

Making Large Language Models Speak Tulu: Structured Prompting for an Extremely Low-Resource Language

We use a structured 5-layer prompt to get GPT, Gemini and Llama to generate grammatically correct Tulu, a low-resource Dravidian language, with no…

Mar 10 • Prathamesh Devadiga

20

2

3