For the past few years i've been on Duolingo & was able to learn Hindi & Spanish (even helped me get state seal of biliteracy)

One of the biggest problems that I see in my generation (Gen Z) is loneliness. I love Duolingo's community, but sometimes (well actually oftentimes), I felt myself struggling to come back.

Online language learning is a skill where dominantly, the entire motivation is connection and the entire daily practice is solitude.

Nobody learns Spanish to know Spanish. They learn it to talk to their partner's family, to not be the silent one at the table, to move somewhere and have a life there. Then they spend 90 days alone with an owl.

I was thinking of a way to connect people around the globe-through the shared passion of learning languages, & creating a competitive social space where people could compete in rapid-fire matches, instantly get feedback, & learn through teaching their own avatar (vs a more passive "lesson-style" learning). They learn on the spot, under pressure, & are put in different situations where they have to maneuver through different absurd scenarios & speak or type the language in a satisfactory way).

The core flow is: Learn (2–5 min interactive segment) → Teach (explain it to your avatar) → Avatar attempts (your score).

Then you take that into ranked 1v1 matches against someone near your level, chess.com style: a blind judge grades both players on speaking and writing, ratings move, there's a global ladder with titles and perks.

The point is to bridge this through competition & internal community-building. Beyond the sport, users can interact with each other & create micro-societies based on their shared interests of the language. People learning Spanish who have gotten an interest in creating foods-stuff like Carne Asada or Gazpacho would join a community (based on their shared interest of the language).

On the technical side I found an interesting mix of problems (that turned out to just be one).

Matchmaking needs P(A beats B) from two ratings. Grading open-ended output is unreliable with rubrics, so instead of scoring one answer I only ever ask "which of these two is better", Comparative Judgment, with scores recovered via Bradley–Terry. Calibrating content difficulty across CEFR/JLPT levels is Item Response Theory, usually a Rasch model.

All three are the same logistic function. Elo is an online gradient step on the Bradley–Terry likelihood; Rasch is Bradley–Terry with one side fixed as "the item." So a 1v1 match is the pairwise comparison the judge needs, the judge's output is the observation the rating update consumes, and item difficulty falls out of the same fit. Ratings are Glicko-2, so a new player carries explicit uncertainty (high RD) and converges in a handful of matches instead of thirty. Users see it as a Fluency Score.

Two Questions: The human one: does competition make you want to practice, or does it just make you anxious about practicing?

If you've shipped LLM-as-judge somewhere the score had real consequences, how did you detect drift you couldn't see from inside the system?

If you think you may benefit/want to use from this, I put together a small waitlist of people who might be interested in using this in the future (I also put a small demo of a little creative writing prompt on the site... see if you can get highest accuracy!)

Thanks for taking the time to read https://thelingo.xyz

Show HN: Ensure your NYC apartment has no roaches, rats, leaks, or problems

NBA suspends Clippers owner Ballmer for 1 year in Kawhi Leonard salary cap probe

Go grandmaster Shin defeats AI KataGo in historic human victory

Photosynthesis – iPhone app combines photos from 2 lenses, out now

You might be wrong about "The Year of Linux on the Desktop"

The Largest Electric Aircraft Just Flew

Climate change past the point of no return worst impacts are now unavoidable

I read all the Star Wars books, and now I understand the prequels [video]

Show HN: Competitive Language Learning (E-Sport)

US gov sides with OpenAI on issue of training LLMs on copyrighted material

Why Google Bid $10M for a Failed Airline's Data

GPS Glitched Across US by as Much as 33 Feet. Scientists Have Never Seen Before

Mr CMO

Hot reload vLLM and sglang configs

FrontierHarness Eval benchmark. Pi is on the Pareto

What Happened to Borland? [video]

The Library of Scott Alexandria

Blast Radius

NameThatUI – What Is This UI Element Called?

Show HN: DeltaCode – Pure AST beat Qwen's zvec-grep on our test (92% vs 48%)

Are Central Banks Moving Out of Dollar Assets?

AIStore at Multi-Petabyte Scale: NVMe versus HDD

Finnish Sauna in Washington DC

The Circular Debt Crisis Nobody Sees Coming [video]

EclipseWise – Predictions for Solar and Lunar Eclipses

Prioritization Is a Lie

Show HN: Redactle LLM Leaderboard

The Marketing Buzzwords That Developers Hate

SF dishwasher job posting offers $50 an hour

It's a Great Time to Be Rich in New York

Show HN: Redactle LLM Leaderboard

Comments

Show HN: Ensure your NYC apartment has no roaches, rats, leaks, or problems

NBA suspends Clippers owner Ballmer for 1 year in Kawhi Leonard salary cap probe

Go grandmaster Shin defeats AI KataGo in historic human victory

Photosynthesis – iPhone app combines photos from 2 lenses, out now

You might be wrong about "The Year of Linux on the Desktop"

The Largest Electric Aircraft Just Flew

Climate change past the point of no return worst impacts are now unavoidable

I read all the Star Wars books, and now I understand the prequels [video]

Show HN: Competitive Language Learning (E-Sport)

US gov sides with OpenAI on issue of training LLMs on copyrighted material

Why Google Bid $10M for a Failed Airline's Data

GPS Glitched Across US by as Much as 33 Feet. Scientists Have Never Seen Before

Mr CMO

Hot reload vLLM and sglang configs

FrontierHarness Eval benchmark. Pi is on the Pareto

What Happened to Borland? [video]

The Library of Scott Alexandria

Blast Radius

NameThatUI – What Is This UI Element Called?

Show HN: DeltaCode – Pure AST beat Qwen's zvec-grep on our test (92% vs 48%)

Are Central Banks Moving Out of Dollar Assets?

AIStore at Multi-Petabyte Scale: NVMe versus HDD

Finnish Sauna in Washington DC

The Circular Debt Crisis Nobody Sees Coming [video]

EclipseWise – Predictions for Solar and Lunar Eclipses

Prioritization Is a Lie

Show HN: Redactle LLM Leaderboard

The Marketing Buzzwords That Developers Hate

SF dishwasher job posting offers $50 an hour

It's a Great Time to Be Rich in New York