What we can learn from Avocado: The unreleased AI Meta’s model - The Next Web
What we can learn from Avocado: The unreleased AI Meta’s model The Next Web
Could not retrieve the full article text.
Read on GNews AI Llama →Sign in to highlight and annotate this article
Conversation starters
Daily AI Digest
Get the top 5 AI stories delivered to your inbox every morning.
More about
modelrelease![[P] PhAIL (phail.ai) – an open benchmark for robot AI on real hardware. Best model: 5% of human throughput, needs help every 4 minutes.](https://d2xsxph8kpxj0f.cloudfront.net/310419663032563854/konzwo8nGf8Z4uZsMefwMr/default-img-graph-nodes-a2pnJLpyKmDnxKWLd5BEAb.webp)
[P] PhAIL (phail.ai) – an open benchmark for robot AI on real hardware. Best model: 5% of human throughput, needs help every 4 minutes.
I spent the last year trying to answer a simple question: how good are VLA models on real commercial tasks? Not demos, not simulation, not success rates on 10 tries. Actual production metrics on real hardware. I couldn't find honest numbers anywhere, so I built a benchmark. Setup: DROID platform, bin-to-bin order picking – one of the most common warehouse and industrial operations. Four models fine-tuned on the same real-robot dataset, evaluated blind (the operator doesn't know which model is running). We measure Units Per Hour (UPH) and Mean Time Between Failures (MTBF) – the metrics operations people actually use. Results (full data with video and telemetry for every run at phail.ai ): Model UPH MTBF OpenPI (pi0.5) 65 4.0 min GR00T 60 3.5 min ACT 44 2.8 min SmolVLA 18 1.2 min Teleop / Fi
How to Replace Your $600/hr Contract Review with a $0.50 AI Analysis
Quick Answer : A law firm just got sanctioned for putting NDAs into ChatGPT. Your firm is next. VoltageGPU’s Confidential Agent Platform runs contract analysis inside Intel TDX enclaves on H200 GPUs for $0.50/analysis — 94% accuracy vs human review. TL;DR : I tested 200 real NDAs. The AI found 47 critical risks my lawyer missed. Time: 62 seconds. Cost: $0.50. Confidential: Hardware-encrypted enclaves (Intel TDX). Why Hardware Encryption Matters ChatGPT processes NDAs on shared GPUs. Your data sits unencrypted in memory. Any hypervisor-level breach exposes it. Intel TDX encrypts data in RAM using hardware. Even we can’t access it. from openai import OpenAI client = OpenAI ( base_url = " https://api.voltagegpu.com/v1/confidential " , api_key = " vgpu_YOUR_KEY " ) response = client . chat . c
Knowledge Map
Connected Articles — Knowledge Graph
This article is connected to other articles through shared AI topics and tags.
More in Models
‘We know your memories and projects need space to grow’: Google just doubled the storage for its AI Pro plan for free while ChatGPT remains limited — here’s how they compare - TechRadar
‘We know your memories and projects need space to grow’: Google just doubled the storage for its AI Pro plan for free while ChatGPT remains limited — here’s how they compare TechRadar
[P] PhAIL (phail.ai) – an open benchmark for robot AI on real hardware. Best model: 5% of human throughput, needs help every 4 minutes.
I spent the last year trying to answer a simple question: how good are VLA models on real commercial tasks? Not demos, not simulation, not success rates on 10 tries. Actual production metrics on real hardware. I couldn't find honest numbers anywhere, so I built a benchmark. Setup: DROID platform, bin-to-bin order picking – one of the most common warehouse and industrial operations. Four models fine-tuned on the same real-robot dataset, evaluated blind (the operator doesn't know which model is running). We measure Units Per Hour (UPH) and Mean Time Between Failures (MTBF) – the metrics operations people actually use. Results (full data with video and telemetry for every run at phail.ai ): Model UPH MTBF OpenPI (pi0.5) 65 4.0 min GR00T 60 3.5 min ACT 44 2.8 min SmolVLA 18 1.2 min Teleop / Fi
Discussion
Sign in to join the discussion
No comments yet — be the first to share your thoughts!