Raw pixels have just outperformed vision encoders across every benchmark. Most multimodal AI systems today rely on stitching together separate components. One encoder processes images, while another generates them. This division causes misalignment and prevents true end-to-end training. A new paper, titled Tuna-2, introduces a breakthrough approach.
@alphasignalai
-
AlphaSignal AI Summarizes AI Breakthroughs
By
–
Paper: https://arxiv.org/abs/2604.24763 Check out http://AlphaSignal.ai for a daily summary of the top AI models, repositories, and research papers. Read by 280,000+ developers.
-
13B Language Model Mimicking 1930
By
–
Researchers trained a 13B language model that thinks it's 1930.
— AlphaSignal AI (@AlphaSignalAI) 9 mai 2026
It speaks like someone from a century ago, with no awareness of computers, World War II, or anything after.
The team behind it, talkie, gathered 260B tokens from public-domain books, newspapers, patents, and… pic.twitter.com/JDbjmh1E3bResearchers trained a 13-billion-parameter language model that believes it is 1930. It communicates like someone from a century ago, unaware of computers, World War II, or any events that followed. The team behind it, Talkie, collected 260 billion tokens from public-domain books, newspapers, and patents.
-
AI tokens wasted on shell command outputs, not code
By
–
90% of the tokens your AI burns aren't from your code.
— AlphaSignal AI (@AlphaSignalAI) 8 mai 2026
Every shell command your agent runs returns raw output.
`git status`, `find`, and `ls` are all read by the LLM in every single run.
A single `playwright test` run dumped enough output to consume 8,100 tokens.
Two… pic.twitter.com/yX1u9lcBo290% of the tokens your AI burns aren't from your code. Every shell command your agent runs returns raw output. `git status`, `find`, and `ls` are all read by the LLM in every single run. A single `playwright test` run dumped enough output to consume 8,100 tokens. Two
-
AlphaSignal.ai’s Mirage Repo Discovery: Daily AI Digest
By
–
Repo: https://
github.com/strukto-ai/mirage
… Check out http://AlphaSignal.ai to get a daily summary of top models, repos, and papers in AI. Read by 280,000+ devs. -
Mirage unifies backends into one virtual filesystem for AI agents
By
–
AI agents just got one virtual filesystem across every backend. Mirage mounts S3, Drive, Slack, Gmail, GitHub, Linear, Notion, and Postgres under a single root. Agents read and write across them using the same bash commands every model already knows. The team rewrote bash
-
AI learns to predict future as neural networks anticipate moving targets
By
–
AI just learned to predict the future, not just react to it.
— AlphaSignal AI (@AlphaSignalAI) 7 mai 2026
A new paper just cracked open how neural networks learn to predict.
Researchers trained recurrent neural networks to chase a moving target.
Some learned to react. Others learned to anticipate.
The difference… pic.twitter.com/zemHjovUqhAI just learned to predict the future, not just react to it. A new paper just cracked open how neural networks learn to predict. Researchers trained recurrent neural networks to chase a moving target. Some learned to react. Others learned to anticipate. The difference
-
Deep Local Research Without OpenAI
By
–
You can now perform deep research on your laptop without OpenAI. Most AI research tools send your queries to the cloud. Your data leaves your device the moment you hit search. Local Deep Research completely reverses this model. It's an open-source assistant that runs entirely on your machine.
-
MIT Researcher Maps Path From Worm to Digital Human Brain
By
–
A former MIT researcher just mapped the path from a worm to a digital human brain. The plan scales from a 302-neuron worm to 86 billion neurons. Three technologies are making this tractable. 1. High-resolution imaging now maps neurons at scale
2. Functional scans capture -
Open-Source Agent Framework Automatically Fixes Its Own Production Bugs
By
–
Someone built an agent that fixes its own bugs on autopilot. Agents fail in production constantly. Hallucinated tool calls, refusal loops, redundant arguments. Fixing these usually means engineers digging through thousands of traces by hand. A new open-source framework