The Work Shaped Object is something people struggle with especially in the world of AI and agents. It might be time to reflect on some of these things and use the excuse we've been given to stop producing unnecessary Work Shaped Objects for shrinking wages.
The Work Shaped Object and AI Agents, General LLMs
Instead of saying "AI can write my email - what do I do now?" it's probably time to question what the next opportunity for work that benefits you could be.
keysmash.dev
Building an LLM means evaluating it over & over as it changes. Tweak a hyperparameter or scale the model up, & every new checkpoint sends you back through the same benchmarking loop. We're releasing olmo-eval, a workbench built for this kind of iterative model development. 🧵
Gemma 4 now runs 2x faster with MTP GGUFs! Run locally on just 6GB RAM. ⚡️ MTP enables Google Gemma 4 run ~1.4–2.2× faster with no accuracy loss. Gemma 4 12B MTP can run at 162 t/s vs. 52 t/s without MTP. 31B reaches 101 t/s. GGUFs + Guide: unsloth.ai/docs/models/...
i’m really loving their work here these models are used for agriculture, forestry, conservation, etc. Basically, if you need some sort of specialized map, this this is going to be very useful
Ai2@ai2.bsky.social · 6mo agoToday we’re releasing OlmoEarth v1.1. It’s 3x cheaper to run than v1 while delivering the same state-of-the-art performance—and fully open. 🧵
Today we're introducing the new Obsidian Community site, the new developer dashboard, and a roadmap of things to come. The future of Obsidian plugins: obsidian.md/blog/future-...
I wasn’t aware that there were non-vibe coded MCPs
Run the Joules@wattasecond.bsky.social · 7mo agoVibe coding an MCP is not recommended
Running opus 4.7 in CC with no system prompt, it's doing well and it just has a much less stressed/anxious vibe
Now that people are catching on to agents authoring agents and skipping MCP for it's variety of context bloat, pollution, and supply chain problems... I'm going to tease the pidgin and wait a few months for people to catch up. GET@google.com/search!q=pidgin>output.txt What's that do?
AI is a bad term because it’s not real intelligence, it’s completely artificial. We need a better name, otherwise we’ll trick people into thinking it’s not artificial
Yes The thing people refuse to want to understand is that Claude is a good "gacha pull" but still required thousands of agents. If you have agents "for free" (and you do, because we have wonderful consumer grade local models) the only thing this costs you is time. :)
Flo 🔶@faz.ms · 7mo agoDon’t have Mythos access yet? Looks like Kimi K2 can find the same 0days with the right harness: aisle.com/blog/ai-cybe...
The big problem with AI is that it can do most things you can imagine already, and people are not accustomed to the infinite probability machine that can do whatever you want.
It's a good time to mention the wonderful world of NPUs and minipcs An AMD Strix Halo will take less power than a mac mini or a desktop PC any day with the NPU load being some < 40W
Asa ⭔@asa.engineer · 7mo agoI do like running my own AI at home, but it's worth acknowledging that my box pulls >100W idle, fights my air conditioner, and has a dismal PUE of probably <~30%, compared to >90% for a datacenter. you should use datacenters more and self-hosting less if you care about energy usage.
Kubernetes is going AI-first! 🚀 We’re proud to support the new Certified Kubernetes AI Conformance program, ensuring AI workloads remain portable and reliable through common, industry-wide standards. Read more: goo.gle/kubernetes-a... #Kubernetes #AI How is your team scaling AI today?
Kubernetes goes AI-First: Unpacking the new AI conformance program
Learn how the Kubernetes AI Conformance program sets open, industry-wide standards for AI workloads via collaboration. Go far, go together!
goo.gle
Small models and community hosted models are the future. Some fully formed adults balk and shriek at the idea because they like to bonsai tree their skills files exclusively using Claude Code, but those people will be relegated to Facebook soon enough :)
Harald Klinke@harald-klinke.de · 7mo ago“Frugal AI” might be the more important story than frontier models. Smaller, open-weight systems → less compute, less energy, more sovereignty and deployable where Big Tech infrastructure doesn’t exist. #AI #DigitalSovereignty #OpenSource #FrugalAI
Gemma 4 is a great set of models you can fine-tune at home
Gemma 4 Lands to Great Reviews
On release, there were a few people suggesting that there was nothing special about Gemma 4. This was missing the point - weight, Gemma 4 outperforms the current landscape of models and has tuning pip...
keysmash.dev
Plot, in case you're losing it: The average person will use AI just like they will use the internet. The current form is not the end Software is now free. It was cheap and enshittified before, outsourcing made it even cheaper. This is why RAM costs so much. Not to make a product but to stop you
This is a small reminder that not long ago the entire idea of a bipedal robot was completely insane. Too much compute & power. Impossible. Until it wasn't. ROBO-ONE started in 2002 with contestants building bipedal robots from hobby servos. Children have built full-body control rigs.
Robo One - Wikipedia
en.wikipedia.org
If an LLM generates an output and there is no one to publish it, does it make a sound?
What's it mean to offer inference publicly? Are you going to jail? Let's walk through our ability to become community inference providers and what that means.
Community Inference - For Good or Bad
There is a surplus of great models that are hostable locally, but not many members of the community have infinite VRAM. Most top out somewhere in 128G. What if we could pool our resources together and...
keysmash.dev
Yikes Lots of Langchain people out there and this one is rough
Hacker & Security News@hacker.at.thenote.app · 7mo agoLangChain, LangGraph Flaws Expose Files, Secrets, Databases in Widely Used AI Frameworks Cybersecurity researchers have disclosed three security vulnerabilities impacting LangChain and LangGraph that, if successfully exploited, could expose filesystem data, environment secre… #hackernews #llm #news
FastFlowLM, an NPU, and Lemonade Server 😎
Qwen3.5-4B — 100% Powered by AMD Ryzen™ AI NPU
YouTube video by FastFlowLM
youtu.be
Some cool news coming out of Pydantic AI. They've merged together the agent gateway and Logfire, which is awesome, but something I'm more interested in is the Capabilities. Define hooks, model settings, and instructions all in one package.
Capabilities - Pydantic AI
GenAI Agent Framework, the Pydantic way
ai.pydantic.dev
anybody seen our new homepage? oxide.computer I had almost nothing to do with it so I can say: it's pretty good. I think it makes clear we sell a big computer that you plug in and turn on. this was a common complaint about the previous site
2006 internet: RSS feeds, forums, communities of hobby blogs, where people post about interior design, fashion, and music for free 2026 internet: sponsored content, paid partnership, buy my merch, try my supplements, invest in my crypto, click link in bio to unlock the secrets of life for $30
