Cleanlab @CleanlabAI

We're thrilled to join forces with @joinHandshake, where we'll be able to scale our team's pioneering work to inflect change with the world's leading AI labs. Hear more from our CEO and Co-founder, @cgnorthcutt, to learn about our next chapter.

Curtis G. Northcutt @cgnorthcutt

4 months ago

News: @joinHandshake acquires @CleanlabAI! This "ten-year old job marketplace" has quietly become a top human data lab for AI--building an AI research org, acquiring top AI talent, and advancing Cleanlab tech and research to lead data foundations for frontier AI. 1 of 4

2 4 22 4K 4

Achieving 20%+ improvement in structured extraction tasks using @DSPyOSS and GEPA Building on a blog post from @CleanlabAI I wanted to see how quickly I could optimize a structured extraction task with DSPy + GEPA In about 3 hours (mostly me getting in the way of claude code): - +22 percentage points over vanilla structured outputs - Ran 4 experiments in total - ~$3 total cost I tested 5 approaches incrementally: • OpenAI Baseline: 32.1% exact match • DSPy Baseline: 39.8% • DSPy + BAML: 42.7% • DSPy + GEPA: 53.8% • DSPy + BAML + GEPA: 54.4%

For anyone who cares about structured output benchmarks as much as I do, here's an early Christmas present 🎁 ! Pretty well thought out from the folks @CleanlabAI. Seems like I'll def be using it to compare LLMs using BAML and DSPy! github.com/cleanlab/struc…

Where Did $37B in Enterprise AI Spending Go? $19B → Applications (51%) $18B → Infrastructure (49%) Our report includes a snapshot of the Enterprise AI ecosystem, mapped across departmental, vertical AI, and infrastructure. Although coding captures more than half of departmental AI spend at $4 billion, the technology is gaining traction across many enterprise departments: IT operations tools ($700M), marketing platforms ($660M), customer success tools ($630 M). AI-native startups are rapidly emerging across every job function, capturing a meaningful share of the $7.3B spent on departmental AI in 2025. mnlo.vc/enterprise-ai-…

Which LLM is better for Structured Outputs / Data Extraction: Gemini-3-Pro or GPT-5? We ran popular benchmarks, but found their "ground truth" is full of errors. To enable reliable benchmarking, we've open-sourced 4 new Structured Outputs benchmarks with *verified* ground-truth

@karanjagtiani04 One example could be: if there is an ambiguous context shift and the agent's original LLM message wrongly assumes something about the context, this can be auto-detected via a low trust score and the auto-revised message can be a follow-up question to clarify instead of assuming

We discovered how to cut the failure rate of any AI agent on Tau²-Bench, the #1 benchmark for customer service AI. Agents often fail in multi-turn, tool-use tasks due to a single bad LLM output (reasoning slip, hallucinated fact, misunderstanding, wrong tool call, etc). We introduce an automated LLM trust scoring + message revision pipeline that mitigates this brittleness and keeps agents on the rails. Benchmarks show that our approach remains effective across all Tau²-Bench domains (Telecom, Retail, Airline) and different LLMs -- cutting agent failure rates up to 50%.

This pipeline can used to automatically make any agent more reliable. Extensive benchmarks here: cleanlab.ai/blog/tau-bench/

👉 Full announcement here: cleanlab.ai/blog/expert-gu…

🚀 New from Cleanlab: Expert Guidance AI agents running multi-step workflows can fail in tiny, trust-breaking ways. Expert Guidance lets teams fix these behaviors with simple human feedback, instantly. ✈️In one airline workflow: 76% → 90% after only 13 guidance entries.

The reality: We’re moving from hype to hardening, building the reliability layer AI needs. 🔍 Read the full Cleanlab report → cleanlab.ai/ai-agents-in-p… 📰 @Computerworld feature → computerworld.com/article/408686…

The “Year of the Agent” just got pushed back. Out of 1,837 enterprise leaders, most are struggling with stack churn + reliability. ⚙️ 70% rebuild every 90 days 😬 Less than 35 % are happy with their infrastructure 🤖 Most “agents” still aren’t really acting yet

🚧 Even the best AI models still hallucinate. OpenAI’s recent paper on Why Language Models Hallucinate shows why this problem persists, especially in domain-specific settings. For teams implementing guardrails, we put together a short walkthrough: youtu.be/i_6fjKgboFg?si…

AI pilots prove intelligence, but AI in production demands reliability. The best teams separate their stack early: 🧠 Core = how AI thinks 🛡️ Reliability = how it stays safe That’s how prototypes become products. 👉cleanlab.ai/blog/emerging-…

@rajistics Love this!

AI agents won’t replace humans. Their real power comes when humans guide it. We just added Expert Answers to our platform: 👩‍🏫 SMEs fix AI mistakes right away 🔁 Fixes are reused across future queries 📈 Accuracy improves, “IDK” drops 10x Full blog: cleanlab.ai/blog/expert-an…

Launching an AI agent without human oversight is basically launching a rocket without mission control 🚀 Cool for a few minutes… until something breaks. 🕹️ It’s not the rocket that makes the mission succeed. It’s the control center. cleanlab.ai/blog/managing-…

📍 Live at @AIconference 2025 in San Francisco! Tomorrow, @cgnorthcutt is sharing practical strategies for building trustworthy customer-facing AI systems, and our team is around all day to connect. 👋 Stop by and geek out with us!

Most AI pilots in financial services never make it to production. The reason is simple: they can’t be trusted. Today, Cleanlab + @CorridorAI are fixing that by combining governance with real-time remediation so AI is finally safe to deploy at scale. 🔗 businesswire.com/news/home/2025…