Haiku 4.5

1. Built an AI flat-finder in a weekend. Indian rental sites are 70% broker spam so I scraped Reddit instead. (www.reddit.com)

   +14 3w haikusonnetanthropic

   Weekend build, ~10 hours. Demo: https://trurent-five.vercel.app/ Problem I was poking at: every major Indian rental site (NoBroker, MagicBricks, 99acres) is infested with brokers even when you filter "direct owner." Reddit actually has hon…

2. I Made LLMs Play Texas Hold’em. The Smallest Model Beat a ~1T Model by Being Too Dumb to Fold (www.reddit.com)

   +71 3w minimaxhaikuanthropic

   Made LLMs play Texas Hold’em against each other. 6 models at the table: a tiny 1.2B running locally on my 16GB MacBook, a couple mid-size ones, and cloud models going up to about 1 trillion parameters.

3. My 1.2B model won 2 out of 5 poker tournaments against models up to 1T params. (www.reddit.com)

   1 3w minimaxhaikuqwen

   I made 6 LLMs play Texas Hold’em against each other. Ran 5 tournaments on my 16GB MacBook.

4. Built a B2B role-play training platform - entirely with Claude (Opus 4.7 backend, Haiku 4.5 for live chat, Claude for design) (www.reddit.com)

   +11 3w haikusonnetopus

   I just launched Socratize (socratize.io) - a rebranded and rebuilt version of FixAI, our original B2C experiment. This time it's B2B-only: teams use it to practice uncomfortable workplace conversations - difficult feedback, client escalati…

5. Anthropic merges consecutive same-role messages, OpenAI doesn't (+4 tokens), anyone token-counted this on open-weight models? (www.reddit.com)

   2 3w haikugpt-5opus+2

   I build context/harness optimization tooling, so provider-side serialization quirks actually matter to me. If you're optimizing over prompts, you need to know exactly what hits the model.

6. How do you reliably override a model's internal temporal bias in production ? (www.reddit.com)

   +11 3w haiku

   I'm building an automated mail generation pipeline using Claude Haiku 4.5 OnPremise but the knowledge cutoff June 2025. This model needs to handle temporal expressions correctly like : next Monday end of the week this month 16 May 16 May 2…

← all threads