AI Agents, Infra Failures, and Dev Tools #7

May 3, 2026 — 3 min read

Today's Letter

DeepSeek released two V4 preview models, DeepSeek-V4-Pro and DeepSeek-V4-Flash, according to Simon Willison's report.
Both are Mixture-of-Experts models with a 1 million token context window; Pro is listed at 1.6T total parameters with 49B active, and Flash at 284B total with 13B active.
The report says the weights use the MIT license, with Hugging Face artifacts around 865GB for Pro and 160GB for Flash.
Pricing is positioned as the main differentiator: Flash is listed at $0.14 per million input tokens and $0.28 output, while Pro is $1.74 input and $3.48 output.
On the cited comparison table, Flash undercuts GPT-5.4 Nano and Gemini 3.1 Flash-Lite on token price, while Pro is placed below larger frontier models from OpenAI, Google, and Anthropic.
DeepSeek attributes the lower cost to efficiency work at long context lengths; in a 1M-token setting, the paper claims materially lower FLOPs and KV cache use versus DeepSeek V3.2.
Self-reported benchmark notes in the paper place DeepSeek-V4-Pro close to top frontier systems, but still short of GPT-5.4 and Gemini 3.1 Pro by an estimated three to six months.
The release is still at single-source stage and not yet independently confirmed across supporting reports.

Source: simonwillison.net

The GitHub repository presents agent-desktop as a native desktop automation CLI for AI agents.
It reportedly controls applications through OS accessibility trees rather than app-specific integrations.
The listing describes structured JSON output and deterministic element references for repeatable automation flows.
Repository metadata shows 357 stars, 16 forks, 2 open issues, and 106 commits at capture time.
The project tree includes crates, npm, docs, tests, and scripts directories, suggesting both core runtime and packaging work.
Support status, platform coverage, and production readiness remain unclear from the available source snapshot.

Source: github.com

Ubuntu and Canonical infrastructure reportedly remained offline for more than 24 hours after a Thursday-morning disruption.
Canonical's status page described a sustained cross-border attack, while most main web properties failed to load.
Reported outages included ubuntu.com, canonical.com, blog.ubuntu.com, developer.ubuntu.com, and security API endpoints.
OS updates from mirror sites reportedly continued, but direct downloads from Ubuntu servers were affected.
The outage followed release of root-level Linux exploit code, limiting Ubuntu's ability to publish security guidance.
A pro-Iran group claimed a DDoS campaign via Beam on Telegram, but affected service scope is still disputed.

Source: arstechnica.com
More: news.google.com

Rancher published K3k, a project for creating and managing isolated K3s clusters inside an existing Kubernetes environment.
The public GitHub repository is rancher/k3k and shows 752 stars, 72 forks, 87 issues, and 17 pull requests at capture time.
Repository structure includes charts/k3k, cli, docs, examples, k3k-kubelet, pkg, scripts, and tests.
Named components visible in the materials include K3k, K3s, k3kcli, and RKE2.
Exposed metadata includes release tag v1.0.2 and API label v1beta1.
The README fragment says the project is intended to run isolated K3s clusters within an existing Kubernetes environment.
License metadata is shown as Apache-2.0.
Runtime details, production readiness, and deployment behavior are not confirmed from the provided sources.

Source: github.com

Jocoletter curates AI, software, and product trends for developers and builders.

#DeepSeek #GitHub #Rancher #Ubuntu