Atlassian recap: Confluence cycle dropped to roughly six weeks

#1 ▶️

Roles aren't converging—they're expanding | Tamar Yehoshua (Atlassian CPO)

Lennys Podcast

#2 𝕏

Aravind Srinivas shared Matterport in a single prompt on Perplexity Computer using the High Effort setting.

#3 ▶️

We Built Grok Bot. Here Are Our 14 Best Bots | Peng Zheng & Lauren Tan

Peter Yang

Also covered by: @Peter Yang

#4 𝕏

NVIDIA AI congratulated AI at Meta on Muse Realtime Avatar and said it worked with Meta’s team to improve the model’s efficiency so the avatar can keep up with users in conversation.

#5 𝕏

Aravind Srinivas recapped running 50–100 workflows on Opus 5.5 over the last few days that previously used Fable 5.1 as the orchestrator, reporting minimal differences. He asked which tasks people still find Fable 5.1 better at than Opus 5.5.

#6 𝕏

Garry Tan recommended the AsideAI browser with MCP for agents facing antibot issues, suggesting it run on a spare laptop or computer kept plugged in. He said it lets agents use the user’s real credentials through a real Chromium browser, calling it a “game changer.”

#7 in

Michael Klingensmith shared that Azure Functions on runtime v3 in a Linux Consumption plan will stop running September 30, 2026, while in-process support and .NET 8 support end November 10, 2026; Linux Consumption tops out at .NET 9, and .NET 10 requires Flex Consumption. Migrating to isolated worker requires code, package, and runtime-setting changes, and Microsoft recommends using a staging slot to manage the two restarts safely.

#8 ▶️

The rise of HI-ICs | Elena Verna (Lovable)

Lennys Podcast

#9 𝕏

Santiago commented that closed frontier models are generally smarter, but self-hosted open models are capable enough for many applications and can be fine-tuned to run quickly at low cost, outperforming leading frontier models on many problems.

#10 𝕏

Sebastian Raschka recapped Ember-1, calling it a great model and an example of his recommended approach to frontier LLM development: start with an existing model and spend a multimillion-dollar budget on post-training.

#11 𝕏

Sebastian Raschka commented that knowledge is generally much more efficient to learn during pre-training, while post-training focuses mostly on skills, behavior, and accessing that knowledge. He cautioned that certainty is difficult without ablation studies at that scale.

#12 𝕏

Lenny Rachitsky shared a conversation with Molly Graham about why her classic “give away your Legos” career advice no longer applies in an AI world. They discuss delegating to AI versus humans, fears of job displacement, responsibilities AI shouldn’t take on, and grief, loneliness, and burnout in tech.

Also covered by: @Lenny Rachitsky

#13 𝕏

Yann LeCun commented that claiming “the domestic robot is here” based on such demos is like saying in 2016 that autonomous cars had arrived after a few minutes of self-driving. He noted that 10 years later, truly autonomous level 5 cars still do not exist.