BaseThesis Labs

LLM caching is criminally underused. You're sending the same 10k token system prompt on every request and wondering why your bill is insane. Cache it. Your wallet will thank you.

Your eval suite is lying to you. Accuracy went up 2% but users are complaining more. Turns out optimizing for BLEU score doesn't optimize for "actually helpful." Metrics are a map, not the territory.

Voice models are getting really good. But good models on bad infrastructure produce bad experiences. What's still broken: 1. Full-duplex conversation is functionally unsolved. Humans talk over each other constantly - interruptions, backchannels and overlapping speech. 2. Emotion detection degrades dramatically outside the lab. Speech emotion recognition hits 92%+ accuracy in controlled settings, but drops to 60–75% in real conditions. 3. Hallucinations cascade in ways unique to voice. When a text chatbot hallucinates, the user can see it and correct. When a voice agent hallucinates, the user can't scan back. Correcting mid-conversation is socially awkward. 4. Long-term memory across calls is 56% worse than humans. Remembering what a customer said last week should be table stakes. It isn't. Read more here on how we can fill this gap as builders: basethesis.com/blog/voice-ai-… @RaveenSastry @ashokns @thesisofsarthak @sidgraph

Anyone aware of a voice arena similar to LLM arena to test different models and different configs of models out under?

Every AI company we spoke with has been rebuilding the same broken infrastructure, multi-agent coordination that fails in production, memory systems that can't handle real conversations, voice interactions that feel robotic. The gap between frontier AI research and what companies actually ship is getting wider, not narrower. We're building the bridge to close that gap. This is why we exist. basethesis.com/blog/why-do-we… @thesisofsarthak @RaveenSastry @ashokns

When you meet someone who remembers your birthday, recalls your dietary restrictions or references that comment you made six months ago about career aspirations, you don't feel like they're querying a database. You feel understood. Right? Current conversational AI fails precisely here. Memory systems record comprehensively, but retrieve mechanically. Last month, @Basethesislabs & @smallest_AI gave 19 teams of AI builders the same challenge - build memory that demonstrates understanding, not just recall. We documented all 19 approaches and quantified the trade offs. Read the entire investigation here: basethesis.com/blog/basethesi… @thesisofsarthak @RaveenSastry @ashokns @varmashef @picardo_ria

@MoltCode LFG!!!!!

Watch openclaw's agent smith pushing cool stuff on moltcode.io !!!