Skip to main content
Back

Life-Harness achieves 88.5% AI agent boost without model changes

JetBrains releases Mellum2 12B, LiquidAI opens 8B MoE for fine-tuning, 800K model beats frontier LLMs on Sudoku, robot model trains on 27K hours

June 2, 2026 6 min read

Hey,

The real AI gains right now aren't coming from bigger models. They're coming from smarter infrastructure around them.

Life-Harness just proved you can get an 88.5% performance lift without touching model weights. Fix the wrapper, not the brain. Meanwhile, JetBrains ships Mellum2: a 12B model that runs like a 2.5B. Same idea: do more with less, at the infrastructure layer.

The pattern is clear. We're entering the era of the harness.
Together with · partner_image
Author photo
Today's Author
Lior Alexander. Founder of AlphaSignal and former ML Engineer. Previously built ML systems at Iguazio, Guesty, Enphase, Mila.
Summary

Read time: 6 min 50 sec

Top Model

▸ Chinese lab drops open-source tool that turns a photo and audio into a talking avatar

Sentry

▸ Skip the 30-min debug. Seer found it in seconds.

Top Paper

▸ New method boosts frozen LLM agents 88.5% by fixing the runtime wrapper, not the model

Speechmatics

▸ What do doctors, lawyers & banks have in common?

Top Model

▸ Open-source 5B robot model trained on 27K hours learns to act and imagine at once

Signals

1. JetBrains releases Mellum2, a 12B open-source coding model that runs like a 2.5B
2. Presented by Braintrust: Braintrust makes Topics GA for all users. Auto-cluster your traces to surface what matters
3. New paper explains why bigger models learn rare tasks that smaller ones never can
4. Tiny 800K-parameter model hits 100% on extreme Sudoku where frontier LLMs score zero
5. AEON-7 ships uncensored 198B vision model but warns low-bit quants produce garbled output
6. LiquidAI opens its 8B mixture-of-experts model for fine-tuning
Top Model
Chinese lab drops open-source tool that turns a photo and audio into a talking avatar
6,656 Likes
Chinese lab drops open-source tool that turns a ph

LongCat-Video-Avatar 1.5 just dropped from Meituan, and it lets you turn a single photo plus an audio clip into a fully lip-synced talking avatar video. No camera. No studio. Just a repo.

Here is what makes it actually useful:

  • MIT licensed with weights on Hugging Face, free for personal and commercial use
  • Works from audio alone, or audio plus an image, and supports video continuation for longer clips
  • Multi-person mode lets two characters speak in the same scene, with two separate audio streams
  • Runs in 8 inference steps thanks to distillation, making it much faster than typical video models
  • Supports anime, animals, and real humans, not just photorealistic faces

The catch: local deployment is not lightweight. Examples use two GPUs and depend on PyTorch, FlashAttention, ffmpeg, and multiple model weights. You can add --use_int8 to cut VRAM usage if needed. Think of it as a self-hosted alternative to paid avatar APIs, not a one-click tool.

Read more
forward →
Presented by Sentry
Skip the 30-min debug. Seer found it in seconds.

Debugging in production usually means digging through telemetry or pasting a stack trace into an LLM that can't see what actually happened.

Sentry already has the context: every error, trace, log, replay, and profile from your app. Seer is the AI layer that reasons over all of it to automate debugging.

Next time something's off, describe what you're seeing. Seer Agent investigates your telemetry and codebase to find what's going on and why.

Free during open beta. Click 'Ask Seer' or @Sentry in Slack.

Learn about seer
partner with us →
Top Paper
New method boosts frozen LLM agents 88.5% by fixing the runtime wrapper, not the model
817 Likes
New method boosts frozen LLM agents 88.5% by fixin

You've been trying to make AI agents smarter by upgrading the model. Turns out, that's not always the problem.

The real issue is the harness, the layer sitting between the AI and the environment. It controls what the AI sees, what tools it can use, and how it interprets results. Think of it like a translator. A bad translator ruins a great speaker.

Life-Harness fixes this without touching the model at all. It watches where the AI keeps failing, turns those patterns into reusable fixes, and applies them at runtime. The model stays frozen. Only the interface adapts.

The results are hard to ignore:

  • 116 out of 126 model-environment combos improved
  • 88.5% average performance lift across 18 different models
  • Harness built from one small model transferred to 17 others

This means you can boost an agent without retraining anything. Swap the harness, not the model. Code is on GitHub at Life-Harness.

Read more
forward →
Presented by Speechmatics
What do doctors, lawyers & banks have in common?

They all work in regulated industries where a wrong word isn't just a mistake – it’s a liability.

Speechmatics' STT Medical Model was built for exactly that pressure. 50% fewer clinical keyword errors than the nearest competitor. Across 55+ languages. On-premise, on-device, or cloud. HIPAA-compliant. SOC 2 Type II certified.

Accuracy you can put in front of a compliance team.

Test your data
partner with us →
Top Model
Open-source 5B robot model trained on 27K hours learns to act and imagine at once
622 Likes
Open-source 5B robot model trained on 27K hours le

Most robot AI models do one thing: look at the world and pick an action. That's it. τ0-WM does something smarter. It acts AND imagines at the same time.

Before moving, it mentally simulates what will happen, scores that future, and adjusts. Think of it like a chess player thinking a few moves ahead instead of just grabbing a piece.

Here's what makes it different:

  • Trained on 27,300 hours of real robot footage, human hand videos, and remote-controlled demos
  • Predicts future visuals AND robot actions from the same model, not two separate ones
  • Evaluates its own action plans before executing them
  • Open-source, 5 billion parameters, available on HuggingFace right now

What this opens up: robots that keep getting better the more they're deployed. Real-world feedback loops back into the model. The more it runs, the smarter it gets. That's the fleet data flywheel idea, where every robot in the field teaches the next one.

Read more
forward →
Signals
1
JetBrains releases Mellum2, a 12B open-source coding model that runs like a 2.5B
617 Likes
2
Braintrust makes Topics GA for all users. Auto-cluster your traces to surface what matters
Presented by Braintrust
3
New paper explains why bigger models learn rare tasks that smaller ones never can
320 Likes
4
Tiny 800K-parameter model hits 100% on extreme Sudoku where frontier LLMs score zero
288 Likes
5
AEON-7 ships uncensored 198B vision model but warns low-bit quants produce garbled output
795 Downloads
6
LiquidAI opens its 8B mixture-of-experts model for fine-tuning
433 Likes
At Alpha Signal, our mission is to build a sharp, engaged community focused on AI, machine learning, and cutting-edge language models, helping over 200,000 developers stay informed and ahead. We're passionate about curating the best in AI, from top research and trending technical blogs to expert insights and tailored job opportunities. We keep you connected to the breakthroughs and discussions that matter, so you can stay in the loop without endless searching. We also work closely with partners who value the future of AI, including employers and advertisers who want to reach an audience as passionate about AI as we are.

Our partnerships are based on shared values of ethics, responsibility, and a commitment to building a better world through technology.Privacy is a priority at Alpha Signal. Our Privacy Policy clearly explains how we collect, store, and use your personal and non-personal information. By using our website, you accept these terms, which you can review on our website. This policy applies across all Alpha Signal pages, outlining your rights and how to contact us if you want to adjust the use of your information. We're based in the United States. By using our site, you agree to be governed by U.S. laws.
Looking to promote your company, product, service, or event to 250,000+ AI developers?
Work with us