Built NotiSift: open-source Android app that sorts each notification into Important / Later / Junk. Offline rules first, AI optional with your own key. Contributors welcome.
https://t.co/j9bMOalF4a
#OpenSource#AndroidDev#Kotlin
The few notifications that matter (a message from a person, a code, a bill due, a delivery arriving) get buried under sales and feed updates. Android only lets you control notifications per app, so you can't say "show Amazon deliveries, hide Amazon deals". I built NotiSift to decide each notification on its own.
Your AI agent is only as smart as what it knows about you.
"Can we make the recital Saturday?" The answer is scattered across your calendar, an email and a note. A personal knowledge graph connects those dots and answers with citations.
Mine runs locally on my Mac Studio. Memory informs; it never authorizes.
What should your agent never know about you?
#AI #AIAgents #KnowledgeGraph #LocalAI
@swyx yeah the human-day estimate from agents is useless. how far off was opus on the ones you actually let run, like 2x or closer to an order of magnitude?
@omarsar0 varying the harness as a scaling axis is interesting. did they find any harness knobs that transferred across tasks, or did each task still want its own weird setup?
@vllm_project fp8 kv as the default for glm-5.x is a fair trade for half the kv memory. is that default going to spread to other models in later releases? would rather know before an upgrade quietly bumps ttft
@emollick the hard to verify part is the real bottleneck imo. a long running task is great until someone has to check the output and theres no cheap way to grade it. who does that review at scale
@akshay_pachaar ok so no grader because it learns from the differences between runs. but if every candidate run fails the same way what is there to learn from? feels like it still needs some signal from outside
@itsnamgyu letting the model say what context it needs in plain text is a neat trick. in a long agent loop what happens if it drops something at step 3 that it needs at step 20, can it ask for it back or is it gone
@jerryjliu0 mining real templates and then varying the data and layout makes sense. how do you check the synthetic docs actually break models the same way the real customer docs do? thats the part id expect to drift
@chamath says software IP is basically worthless now. @Benioff : "Our API is the UI."
Headless = agents do the work, no screen. The value moves to the connected ecosystem, and you watch token cost per task.
What would you still pay for? https://t.co/SHmcWrlMBp #AgenticAI#MCP
@pauliusztin_@zenml_io caching tool outputs makes sense. but if the model samples a different path on replay it'll call a tool that isnt in the cache. do you fail the replay there or fall through to live?
What about legal privacy etc
@Thom_Wolf shorter outputs is the line i care about too. for agents the bill is tokens per finished task, not price per token. has anyone measured tokens per solved task vs the old flash yet
Phone reminders get swiped away, so my Grok Bot team puts them on the kitchen screen.
One bot per job: the calendar bot casts a card to our Nest Hub 30 min before each event and says it by name. No calendar edits without my OK.
Full breakdown + video 👇
https://t.co/MKuxOaVGCN
#GrokBot #AIAgents
Tesla Roadster is finally here, 9 years after the prototype 🚀
⚡ 0-60 in under 1 second
🌀 Fan suction holds it upside down
🛸 Rocket trim can fly (not road legal)
🏁 Road, race & air tiers
📅 2028 target, price TBA
Drive it or fly it? 👇
#Tesla#TeslaRoadster#ElonMusk#EV
@river_ai_inc@NVIDIAAI what does the eval score, exact match or execution results? in my experience schema context and past queries move text-to-sql more than the base model, so curious how much the recipe vs the model did here
@AverageAiBro this is the part people skip. prompt guardrails aren't access control. does it filter columns too or just whole tables? payroll usually hides in one column of a "safe" table