Join /aichannel https://cura.network/aichannel/details?a

2 replies

1 recast

33 reactions

GLM-4.7: the new open source star! The Takeaway: 👉 GLM-4.7 boosts coding and complex reasoning with enhanced planning mechanisms and consistent multi-turn thinking. 👉 The model shows substantial benchmark gains over GLM-4.6, especially in agentic and tool-use scenarios. 👉 It's fully open source and accessible via APIs, making advanced AI capabilities broadly available to developers. 👉 Practical use cases include smarter coding agents, deeper context understanding, and more robust long-form tasks. The AI world just got a major upgrade: GLM-4.7, Z. ai's newest flagship model, has officially launched with a clear focus on stronger coding, reasoning, and multi-step task performance. At its core, GLM-4.7 builds on the impressive foundation of earlier GLM models by enhancing interleaved thinking, a mechanism that lets the system plan ahead before responding, and adding preserved and turn-level thinking for better consistency over long, complex conversations. These upgrades make it much better at handling tasks that require deep logical steps, like debugging code, maintaining context across turns, or orchestrating actions with external tools. But here's what really makes GLM-4.7 stand out for the AI community: It significantly improves real-world coding metrics and reasoning benchmarks compared to its predecessor, while remaining fully open-source and accessible through APIs and platforms like OpenRouter and PPQai (https://ppq.ai/invite/7852e16b) Imagine an AI that not only writes cleaner, more efficient code, but also keeps the "thought process" behind the scenes more stable and reliable - that's what GLM-4.7 aims to deliver. GLM-4.7 pushes the envelope for open-source models in coding and reasoning tasks, offering powerful capabilities without API lock-in. It strengthens the bridge between research and real-world developer workflows in the AI ecosystem. Sources: https://z.ai/blog/glm-4.7 https://docs.z.ai/guides/llm/glm-4.7

1 reply

0 recast

1 reaction

50+ Expert Predictions: Ways to Drive Agentic AI, Data Governance, and Security in 2026 https://drive.starcio.com/2025/12/predictions-agentic-ai-data-governance-security-2026/

0 reply

0 recast

0 reaction

I'm missing couple of my most important chat threads in the ChatGPT app and ChatGPT web. Has anyone else experienced this and resolved it?

1 reply

0 recast

2 reactions

FunctionGemma, Google puts an AI agent directly on your phone Google has quietly released FunctionGemma, a compact AI model that runs entirely on-device and can turn voice commands into real system actions. No internet, no cloud, no data leaving your phone. What it is: • 270M-parameter model optimized for mobile • Fully offline and private by design • Converts natural language into executable phone actions • No servers, no latency from network calls How it works: • You speak a command • The model parses intent locally • Translates it into structured system functions • The phone executes instantly Example: “Add John to contacts, number 555-1234” → contact saved, zero cloud involvement. Why it’s different: • Traditional assistants rely on cloud round-trips • FunctionGemma runs at ~0.3s latency on-device • Privacy by default, nothing leaves the phone • Works even with no signal Performance snapshot: • 270M parameters (≈6,600× smaller than GPT-4) • 126 tokens/sec • ~85% accuracy after fine-tuning (up from ~58%) • ~550 MB RAM usage • ~288 MB file size • ~0.75% battery drain for ~25 interactions What it can do: • Set alarms and reminders • Create calendar events • Control smart-home devices • Navigate locally • Log simple activities All processed locally. All instant. Architecture insight Google frames this as a “traffic controller” model: • Simple, sensitive tasks → handled on-device • Complex tasks → optionally routed to cloud AI Limitations • No multi-step task chaining yet • Struggles with indirect phrasing • 15% error rate without tuning • Best results require app-specific fine-tuning This isn’t about one assistant. It’s a shift in philosophy: small, specialized models doing real work locally can outperform larger general models, faster, cheaper, and privately. https://huggingface.co/google/functiongemma-270m-it

0 reply

0 recast

0 reaction

It's a shame that @grok is really bad with abstract images and ideas. It is excellent when it comes to people and real-life scenes though.

1 reply

3 recasts

10 reactions

Cars will become the 4th space ... I'd still keep the 3rd space for the community and coffee shop 😉 https://x.com/signulll/status/2002222118348419328?s=20

0 reply

0 recast

2 reactions

"A Guide to Local Coding Models" https://www.aiforswes.com/p/you-dont-need-to-spend-100mo-on-claude

0 reply

0 recast

0 reaction

A superpower is building in public, then admitting you were kind of wrong, but going on anyway! "But do I want someone reading this to immediately drop their coding subscription and buy a maxed out MacBook Pro? No, and for that reason I need to correct my hypothesis from ‘Yes, with caveats’ to ‘No’." Personally, I think the cloud models are pretty awesome, but if I was say in Hong Kong, I would definitely consider investing more in local models. Would be nice to cover Mistral as well, but its a good base article. https://www.aiforswes.com/p/you-dont-need-to-spend-100mo-on-claude

1 reply

4 recasts

14 reactions

Anthropic Bloom: an open source tool for automated behavioral evaluations https://www.anthropic.com/research/bloom

0 reply

0 recast

2 reactions

the kinds of idiotically stupid these language models are, is always hilarious to me

1 reply

0 recast

2 reactions

Claude in Chrome. Make sure wallets are elsewhere. I've casted about this before for other AI-driven browsers, and I like the Claude strategy (because they have to outperform Google's native strategy), but I am still weary of it, as you should be. They warn you too, which is a Good Thing™. https://claude.com/chrome

0 reply

2 recasts

7 reactions

How to build a defensive AI security agent with RAG https://boringappsec.substack.com/p/how-to-build-a-defensive-ai-security

0 reply

1 recast

2 reactions

I’m building a multi-state agent where only certain instructions are relevant depending on how far along the user is in the conversation so Skills looks like an ideal design pattern that I’m quite excited to try out. And I learned about it 2 days ago from @jachian — thanks!! https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills

0 reply

0 recast

3 reactions

Send 🎁 to ChatGPT. Get a nice little video gift generated by Sora. They tell you to get the app as well, but if you're not in the USA App Store, it is still not available. Time for a worldwide release?

2 replies

0 recast

3 reactions