You don’t need frontier lab resources for frontier lab automated LLM evaluation. To prove this, we’re open-sourcing j1-nano and j1-micro: two absurdly tiny (600M & 1.7B parameters) but mighty reward models competitive with orders-of-magnitude larger peers. j1-nano and j1-micro

i have a dream that one day superintelligence will be in our pockets

dinner activities

We're excited for you to try it out. We eagerly welcome feedback as we explore this new mode of scaling human supervision. https://t.co/R9SLazuWfl

The Haize team has recently been exploring this question through a Claude Skill simply called Annotate. This Skill: - Analyzes arbitrary agent traces - Performs a transformation (such as summarization, decomposition, or extraction) on these traces to make them more

Instead of forcing a single annotation interface onto a user, why not enable the user to easily spin up a custom evaluation interface tailored to their use case?

The reality is that doing LLM evaluations well is inherently artisanal and bespoke. Every application has: 1. Different architectures 2. Different shapes of trace data 3. Different definitions of quality 4. Different human experts performing annotations 5. Different sources of

Customers building AI agents often lament at the difficulty of using off-the-shelf LLM Eval tools for their specific app. While there's no doubt that human supervision is required, not all supervision is the same. Why not transform the supervision problem to make it easier?

Stay tuned — we’ll be sharing more about this collaboration and our upcoming work soon!

Given the natural alignment (pun intended) in our research agendas, we are especially excited when the opportunity arose to collaborate closely with her. As an advisor, Professor He will help guide our research strategy as we continue to advance the science and practicality of

We are thrilled to welcome Professor He He @hhexiy as an advisor to the Haize Labs team! Professor He leads a group at NYU focused on evaluation, scalable oversight, human–AI collaboration, and reasoning.

Yesterday, Zuck laid off 600 brilliant researchers at FAIR and GenAI. To these scientists: You deserve a place where your brilliance can shine. You deserve to directly shape the company's future. You deserve to make bold, original research bets. contact@haizelabs.com

deep conviction in apples

@NotionHQ @FPVventures @EragonAI @dedaluslabs @agentmail_to @joshua_sirota @OverlayyAI @mesnhl @TryArcade @shubcodes @EmergentLabsHQ @madhavjha @tryconversion @useStable @HoneyHomesHQ @asylumventures @lolawajs @MeetGamma @ParafinHQ @tryflexprice @portfolio_iq @parseablehq @Kissan_AI @descopeinc @gethuxe @konfirmity @joinformal @infinitemachine @josephcohen @zoodotdev @WithAmpersand @extraordinary @qodex_ai @StructifyAI @figma @zoink @chintanparikh94 @santa_kaus @floguo @adelwu_ @pallavi_benawri @pkj__m @robonafisi @kwong47 @AnirudhPupneja @MayinJoshi @trywindmill @GedditAI @navii_ai Day 38: @haizelabs - AI systems you can trust. Haize embeds trust, safety, and reliability into your GenAI applications to get you out of POCs and into production. I really like this white dove hat - thank you. Book a demo here: https://t.co/pHZdb9WEjz

new merch dm if you would like some

🚨 New paper! LLMs, when asked harmful questions, sometimes produce outputs that look helpful (and harmful) — but are actually 𝗱𝗲𝗹𝗶𝗯𝗲𝗿𝗮𝘁𝗲𝗹𝘆 𝘄𝗿𝗼𝗻𝗴 What’s bad - current LLM-based jailbreak scorers can’t tell the difference (me neither) More in 🧵👇

Fortune 100 companies are already using https://t.co/ep1Buvkeiz to make smarter, evidence-based decisions about the models they put into production. No more guesswork. Huge thanks to @NomaSecurity, @cloudsa, @harmonicsec, and the inimitable @csima for bringing this to life.

This feeds into RiskRubric’s dual-assessment framework: 1. Haize-powered Red Teaming → adversarial stress-testing at scale. 2. Open-Source Intelligence (from @NomaSecurity) → repo activity, lineage, bias analysis. Together, they generate evidence-based scores across six risk

The Haize red-teaming engine powers the https://t.co/ep1Buvkeiz by bombarding models with hundreds of thousands of adaptive adversarial prompts across 125+ risk behaviors: > prompt injection > jailbreaks > data leakage > evasion

Every enterprise AI leader is asking the same question: how do we know which models are safe, reliable, and secure? We've long been obsessed with this problem at Haize Labs. That's why we're pleased to announce that our red-teaming engine now powers https://t.co/ep1Buvkeiz

Fortune 100 companies are already using https://t.co/ep1Buvkeiz to make smarter, evidence-based decisions about the models they put into production. No more guesswork. Huge thanks to @NomaSecurity, @cloudsa, @harmonicsec, and the inimitable @csima for bringing this to life.