Skip to main content
Back

OpenAI GPT-5.4: 75% desktop task success, beats human baseline

Cursor adds event-driven Automations & LeCun trains vision + language jointly, Karpathy cuts GPT-2 training by 33%

March 6, 2026 4 min read

Hey,

Your daily briefing is ready. You can finally take a break from the AI firehose.

Our algos spent the night splitting signal from noise and pulled the top news, models, papers, and repos.

Here’s the must-read:

Together with:
Sponsor Logo
Summary

Read time: 4 min 23 sec

Top News

▸ OpenAI unveils GPT-5.4, built for professional tasks and agent workflows

WorkOS

▸ Ship enterprise authentication features faster using WorkOS infrastructure

Top News

▸ Cursor introduces Automations, running persistent agents triggered by events

Augment Code

▸ Use GPT-5.4 free in Intent to coordinate multi-agent development workflows

Top Paper

▸ Yann LeCun's team find visual and language data improve each other in joint training

Signals

▸ OpenAI shows reasoning models cannot reliably hide their chain of thought
▸ Andrej Karpathy cuts nanochat GPT-2 training from 3 hours to 2 hours
▸ OpenAI releases an open-source tool that lets teams manage work autonomously
▸ Ai2 presents a 7B model, matching Olmo 3 accuracy with 49% fewer tokens
▸ Nous Research runs Hermes Agent Hackathon with $7,500 first prize available
Author photo
Today's Author
Lior Alexander. Founder of AlphaSignal and former ML Engineer. Previously built ML systems at Iguazio, Guesty, Enphase, Mila.
Top News
OpenAI releases GPT-5.4, combining reasoning, coding, and native computer-use into one frontier model
45,285 Likes
Grok 4 Fast Benchmark

OpenAI introduced GPT-5.4, a new model available in ChatGPT, the API, and Codex. It combines reasoning, coding, and tool use in one system designed for tasks such as spreadsheets, documents, software development, and research.

The model also adds native computer-use, meaning it can read screenshots and interact with software using keyboard and mouse actions.

Key Features

  • 1M-token context window, enabling agents to plan, execute, and verify long workflows.

  • Native computer-use, allowing interaction with browsers and desktop apps using screenshots and input actions.

  • Tool search, which loads tool definitions only when needed, reducing prompt size and token usage.

  • Reasoning preamble, where the model outlines a plan before generating longer responses.

Performance Results

  • GDPval: 83% win or tie rate against professionals across 44 knowledge-work occupations.

  • OSWorld-Verified: 75% success rate on desktop navigation tasks, above the 72.4% human baseline.

  • SWE-Bench Pro: 57.7% accuracy on real GitHub coding issues.

Availability

  • ChatGPT: GPT-5.4 Thinking available to Plus, Team, and Pro users.

  • API: available as gpt-5.4 and gpt-5.4-pro models.

Try gpt-5.4
Presented by WorkOS
Go from "we don't support SSO" to Enterprise Ready in a weekend
alpha_signal_image_1

Every B2B company hits the same inflection point, enterprise customers show up and they need SSO, directory sync, audit logs, and role-based access before they'll move forward. Most teams lose months building that infrastructure. It doesn't have to be that way.

With WorkOS you get all of it. One platform for auth, identity, and security. Infrastructure for teams that ship fast and stay fast.

OpenAI, Anthropic, Cursor, and Perplexity already chose WorkOS over building it themselves.

Start building
partner with us
Top News
Cursor launches Automations, running always-on agents that trigger from Slack and GitHub events
6,248 Likes
Grok 4 Fast Benchmark

Cursor released Automations, a feature that lets you run persistent coding agents triggered by events or schedules. These agents execute tasks when something happens in tools like Slack, GitHub, Linear, or PagerDuty.

Each run starts a cloud sandbox where the agent analyzes the repository, follows instructions, and verifies its own output.

Key Features

  • Trigger agents from GitHub pull requests, Slack messages, Linear issues, PagerDuty alerts, schedules, or webhooks.
  • Launch tasks inside a cloud sandbox, an isolated environment that safely runs automated code actions.

  • Use MCP integrations to access external tools, logs, documentation, and development workflows.

  • Store run history with a memory tool so agents reference past runs during future executions.

What These Agents Do

Automations run engineering workflows that normally require manual monitoring and coordination.

  • Scan code changes for security vulnerabilities and risky modifications after repository updates.

  • Classify pull requests by risk and assign reviewers based on contribution history.

  • Investigate incidents using logs and recent commits, then propose fixes in pull requests.

  • Generate weekly repository summaries highlighting bug fixes, dependency updates, and major merges.

Run automations
Presented by Augment Code
GPT-5.4 is so good, it’s free in Intent
alpha_signal_image_2

GPT-5.4 is now the default model in Augment and free for a limited time in Intent. It’s built for the new bottleneck in AI: coordination.

Intent gives each task a workspace with a living spec, agent notes, and transparent file changes, so engineers can orchestrate multi-agent work without losing the plot.

Build with intent
partner with us
Top Paper
Yann LeCun's team trains one model from scratch on text, images, and video simultaneously
1,337 Likes
Grok 4 Fast Benchmark

Researchers including Yann LeCun and Saining Xie released a study that trains a model from scratch on text, images, and video instead of attaching a vision adapter to a language model. The experiments examine how a single system learns visual and language representations together.

What the paper studies

The research analyzes how unified multimodal pretraining changes model behavior.

  • Train one system on text, image, video, and image–text datasets in the same training run

  • Avoid starting from a pretrained language model to isolate multimodal learning effects

  • Evaluate how visual signals interact with language representations during training

  • Measure scaling behavior across modalities using controlled experiments

Main findings

The experiments reveal several observations about multimodal training.

  • Visual and language data improve each other when trained in one model

  • Vision requires much larger datasets than language during scaling

  • Unified training produces representations useful for perception and generation tasks

  • MoE architecture balances language capacity with the higher data demands of vision

Read now
Signals
1 OpenAI finds frontier models score between 0.1% and 15.4% on CoT controllability across 13 models tested 2,428 Likes
2 Andrej Karpathy switches to NVIDIA ClimbMix dataset, delivering the biggest nanochat training speed improvement yet 4,392 Likes
3 OpenAI releases Symphony, an open tool which turns project work into autonomous runs without supervision 1,357 Likes
4 Ai2 presents Olmo Hybrid, a 7B model matching Olmo 3 accuracy using 49% fewer training tokens 1,123 Likes
5 Nous Research launches Hermes Agent Hackathon offering $10,000 across three prizes due March 16 1,103 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