Skip to main content
Back

1980s energy model beats modern AI loops by 1,000x in speed

Anthropic upgrades Claude Code Projects, DeepSeek opens API features, Qwen4-Max cuts costs 60%, 18MB TTS model runs on CPU

November 16, 2026 6 min read

Hey,

Claude Code got custom themes and smarter project coordination this week. Fun, but almost decorative next to the real story.

CRAM just showed that a 1985 physics idea outperforms modern reasoning loops by 1,000x in training speed. We keep building bigger, and a 40-year-old energy function quietly solves Sudoku in 42ms on one GPU.

The pattern: the best ideas aren't always the newest ones.
In Partnership 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 9 sec

Top News

▸ Claude Code desktop app now supports custom themes with animated backgrounds

WorkOS

▸ Your AI agents shouldn’t have all your permissions

Top News

▸ Anthropic upgrades Claude Code Projects with smarter coordination and thread sharing

Top News

▸ DeepSeek opens web search, file parsing, and voice to API developers

Signals

1. Qwen releases Qwen4-Max with 60% lower cost and 1M-token context
2. Buoyant.io: Kinference: open-source inference serving for platform teams
3. Open-source tool reverse engineers any codebase into plain business workflows
4. 1980s energy model trains reasoning tasks 1,000x faster than looped models
5. Meta study of 200+ models finds vision encoders need parameters, not pretraining
6. Open-source 18 MB text-to-speech model runs on CPU, never trained on real voices
Our Events
From Vulnerable Dependencies to Tested Pull Requests with Snyk
Dec 9, 10a PT · Online · Register now
When AI Agents Ship Faster Than Security Can Review
Dec 17, 10a PT · Online · Register now
Top News
Claude Code desktop app now supports custom themes with animated backgrounds
23,490 Likes
Claude Code desktop app now supports custom themes

Claude Code just got a glow-up. You can now build custom themes for the Claude Code desktop app using a feature called mods, which are small plugins that change how the app looks and behaves.

Want a dark purple background with animated particles? Just describe it to Claude and it builds the theme for you. No design skills needed.

Here is what you can do right now:

  • Change colors across the entire UI
  • Add animated backgrounds
  • Use the theme-examples plugin to get started fast with pre-built scenes
  • Write your own theme from scratch using the official docs

To try it, update to the latest desktop version. Themes currently work in local sessions only.

The fun part: you do not need to write a single line of CSS. Just tell Claude what vibe you want. It handles the rest. Your terminal, your aesthetic.

Read more
forward →
Presented by WorkOS
Your AI agents shouldn’t have all your permissions

Most agents run with the full access of whoever triggered them. WorkOS Airlock is the authorization layer for agentic software. You write policies, and Airlock checks every agent call at runtime, then allows it, denies it or sends it to a person for approval.

Agents never see credentials. A risky action, like a refund over your limit, goes to a teammate to approve in Slack. Airlock works with coding agents, custom harnesses and any API or MCP server.

Learn more
partner with us →
Top News
Anthropic upgrades Claude Code Projects with smarter coordination and thread sharing
21,615 Likes
Anthropic upgrades Claude Code Projects with smart

Claude Code just got a solid upgrade to its Projects feature, and it changes how you manage complex, multi-part work.

Projects let Claude run multiple tasks at the same time in separate threads, like having several conversations going in parallel. The coordinator, the part that oversees everything, now speaks up more. It answers your questions directly in the project chat and gives you live status updates on what's running.

  • Share threads as read-only links with anyone signed into Claude
  • Set reminders on any thread so it hides until you need it, then Claude surfaces it automatically
  • /btw spins off a new thread mid-task so your current one stays focused
  • /fork branches a thread so you can try a different approach without losing your original
  • Artifact thumbnails now show inside each thread for faster navigation

Projects is in public beta on Pro and Max plans. Join the waitlist if you don't have access yet.

Read more
forward →
Top News
DeepSeek opens web search, file parsing, and voice to API developers
14,042 Likes
DeepSeek opens web search, file parsing, and voice

DeepSeek just gave its V4.1-Flash model a serious upgrade. The tools powering the DeepSeek consumer app are now available through the API. No more stitching together third-party services.

Before this, adding web search or PDF parsing to your AI app meant calling external APIs yourself and manually assembling the context. Until now, incorporating web search, PDF parsing, or speech recognition required setting up your own infrastructure. That changes now.

Here is what you can plug in with a few parameters:

  • Web Search: the model searches the web and returns cited answers automatically
  • Deep Thinking Search: searches and reads pages during multi-step reasoning
  • File Input: drop in PDFs, Office files, or images and the model reads them
  • URL Reading: paste a link in a message and the model fetches and reads it
  • Safety Checks: screens inputs and outputs, flags risky content automatically

Enable search with deepseek_web_search and URL reading with deepseek_read_url. Tool usage is billed on top of normal token prices. Public apps must have safety checks on.

Read more
forward →
Signals
1
Qwen releases Qwen4-Max with 60% lower cost and 1M-token context
7,306 Likes
2
Kinference: open-source inference serving for platform teams
Presented by Buoyant.io
3
Open-source tool reverse engineers any codebase into plain business workflows
1,805 Stars
4
1980s energy model trains reasoning tasks 1,000x faster than looped models
2,070 Likes
5
Meta study of 200+ models finds vision encoders need parameters, not pretraining
855 Likes
6
Open-source 18 MB text-to-speech model runs on CPU, never trained on real voices
5,676 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 300,000 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 300,000+ AI developers?
Work with us