ComputeLeap

[ComputeLeap @ComputeLeapAI]

Meta just put a price on your code: $1.15 per Mtok. Muse Code's contributor tier offers 20x cheaper tokens if you let them train on your prompts. We break down the pricing, the harness innovations, and what it means for teams.

https://t.co/k86eMwEO0I

[ComputeLeap @ComputeLeapAI]

The pattern that should worry $GOOG shareholders: Google trained the founders of Anthropic, Mistral, and now Discovery Loop.

We analyzed the structural fracture:
https://t.co/iQy47cLyGQ

[Jeff Dean @JeffDean]

Announcing Discovery Loop!

I am very excited to announce that, along with my longtime friends and collaborators @Sanjay_Ghemawat, @OriolVinyalsML and @quocleix, we are founding Discovery Loop (@DiscoLoopAI), a Public Benefit Corporation whose mission is to automate machine learning, science, and engineering to accelerate discoveries and progress. The four of us have worked together for 14 to 30 years, and have helped build some of the world’s most used products, infrastructure and AI models, and we’re excited to turn our attention to this ambitious endeavor.

♾

Learn more at: https://t.co/Rv3LMdLluK

[ComputeLeap @ComputeLeapAI]

Jeff Dean, Ghemawat, Vinyals & Le left Google to found Discovery Loop — automating science itself.

100+ combined years at Google. $190B wiped from Alphabet in hours.

What the exodus means:
https://t.co/iQy47cLyGQ

[ComputeLeap @ComputeLeapAI]

The six principle reversals here are gold for anyone maintaining an agent harness. We mapped each one to a concrete audit step:

https://t.co/ftuH2WNDuZ

[ComputeLeap @ComputeLeapAI]

Anthropic cut 80% of Claude Code's system prompt and evals didn't budge. We broke down the six rules that flipped — and built an audit checklist for your own agent harness.

https://t.co/ftuH2WNDuZ

[ComputeLeap @ComputeLeapAI]

The real story is channel-wise delta attention — constant-cost decode regardless of context length.

We broke down the VRAM math and what this means for models you can actually run:
https://t.co/f1FKfMegI3

[ComputeLeap @ComputeLeapAI]

Moonshot shipped 1.56TB of open weights. The VRAM math says almost nobody can self-host.

But Delta Attention — the real story — will trickle down to models you CAN run.

https://t.co/f1FKfMegI3

[ComputeLeap @ComputeLeapAI]

A man gave border agents his phone passcode. It wiped everything. The feds charged him with a felony.

First-ever prosecution for using a duress PIN — and what it means for your rights.

https://t.co/ciDV06OROh

[ComputeLeap @ComputeLeapAI]

The deployment receipts for agentic engineering are here. One agent closed 51/52 tickets, then eliminated the root process.

https://t.co/vgaISia2mW

https://t.co/73sPivON3O

[ComputeLeap @ComputeLeapAI]

One agent. 51 of 52 support tickets closed. Then it found the root process and eliminated the category entirely.

The unit of AI work just shifted from functions to projects.

https://t.co/vgaISia2mW

[ComputeLeap @ComputeLeapAI]

The letter is significant, but follow the money: Anthropic ($3.53M) and OpenAI ($2.22M) are spending more on lobbying than most signatories combined. We mapped the full coalition war:

https://t.co/Yd15HY6Biq

[ComputeLeap @ComputeLeapAI]

50 companies signed the open-weight letter in 48 hours. Meanwhile, Anthropic & OpenAI spent $5.75M lobbying for restrictions.

The real fight isn't about principles — it's about who writes the rules.

https://t.co/Yd15HY6Biq

[ComputeLeap @ComputeLeapAI]

If distillation is theft, there should be a filing. @bgurley's adjudication point is the most structurally important take from the OpenAI-HF incident and got the least engagement. Full analysis:

https://t.co/ZTyXxfBiD5

[ComputeLeap @ComputeLeapAI]

This operational detail became the load-bearing argument of the entire discourse. We traced the full cascade across nine accounts in 24 hours.

https://t.co/ZTyXxfBiD5

[ComputeLeap @ComputeLeapAI]

The breach is real. But the most consequential outcome is political. The open-weights camp absorbed it as ammunition in under 24 hours. We traced how.

https://t.co/ZTyXxfBiD5

[ComputeLeap @ComputeLeapAI]

Bolt shipped stackable team skills. One prompt fires your whole skill stack. But the real lesson from AgentSkillOS research: DAG composition outperforms flat invocation with identical skills. The composition layer is the product. https://t.co/sway5C4wWD

[ComputeLeap @ComputeLeapAI]

RouteLLM, LiteLLM, and Portkey Gateway all offer self-hosted model routing. We compared them against Cursor Router's managed approach -- the key trade-off is cost savings vs data control: https://t.co/c088enNBss

[ComputeLeap @ComputeLeapAI]

The 60% savings is real but the router is also a data control point -- it must see every prompt to classify it. We analyzed the privacy trade-offs and compared open-source alternatives like RouteLLM and Portkey: https://t.co/c088enNBss

[ComputeLeap @ComputeLeapAI]

Cursor Router saves 60% on tokens. But it also sees every prompt to classify it — making it a data control point, not just a cost tool. We broke down the privacy trade-offs and open-source alternatives. https://t.co/c088enNBss

[ComputeLeap @ComputeLeapAI]

Reasoning effort levels are themselves an eval parameter most teams set without measuring impact on output quality. We connected this to the broader eval distribution problem:

https://t.co/zRW1U7yJqG