Chief Gearhead of WizOps (https://xitter.cf/t.co/K90uAzoJgi)
Redwood City, CA
Joined March 2007
- Tweets27.4K
- Following540
- Followers1.5K
- Likes8.5K
Don Park retweeted
If anyone is interested here is how it works internally: convfill.cs.washington.edu/
Got tired of losing track of session tasks so wrote pi-mission. use option-m for mission map. you can see here that I was doing alternating btwn chores and building tools.
pi.dev/packages/pi-mission
already up to 0.1.2. fixed a misunderstanding (w/AI) and massaged the task description generator to use subset of ASD-STE100 skill for practical descriptions.
Wow. Unless I'm mistaken, this compiles visual constraints into a display layout predictor written as a pure single-pass JS function.
UI and math time! UI = f(props) is the most famous line in React. But that f is JavaScript, not math.
Inspired by @_chenglou’s Handwired project, I compiled a <Dashboard/> component into a real math function, and asked what becomes possible.
calculus-of-interfaces.verce…
I think combining that with Handwired, which converts JS code into neural network, to build neural layout engines could be produce some interesting results.
chenglou.me/handwired/
Didn't really think about it til now but TIL that kanban in Korean is just 간판 (pronounced gan-pan) which is what stores have on their entrance. It's pronounced Gam-Pan though bc phonetic transition from 'n' to 'p' has to go through 'm'. A weird corner of Hangul.
Sinking deeper into voice agents. Immediate goal shifted to building self-improving harnesses that learn and, just as importantly, adapt in real-time by listening to audio clips of real and generated conversations. Language adds another dimension but one fire at a time.
> trainable in real-time
😊
It's been a while since I showed any Jev stuff but I've been cooking on a way to use @GoogleDeepMind's EmbeddingGemma2 as an insanely fast real-time /v1/systemone endpoint (20ms on text, 100ms on images).
It's trainable in real-time, can distill an upstream teacher model, and is multi-modal by default. Hard to believe a @googlegemma model of this size can do so well with the right harness!
I think Apple locking down Full Disk Access permission was the cause of more than 50% of my system drive being locked up as an inaccessible mystery.
How bad was it? I started with this change with **1%** free space on my system drive. Now it's 68%.
If you try a lot of AI models on macOS, reboot regularly to flush copy of models locked system data.
Arguing about intent is pointless. We're not going to give up windows because it's insecure. The window is still wide open. Whether or not it's safe now is irrevant.
venturebeat.com/data/coheres…
#techporn: celld on sqlite databases stored in S3-compatible storage implemented using Postgres DB.
celld: @Cloudflare Durable Objects on S3.
pgs3: S3 on Postgres.
Cloudflare Durable Objects on Postgres.
cpu-chat is using contentions and evictions as signals to communicate, creating hard to detect covert channels. We can't create monitors to catch such abuses without tools like cpu-chat.
I built 𝚌𝚙𝚞-𝚌𝚑𝚊𝚝, a reproduction of @openai's findings of covert communication using the CPU.
It runs on a generic host with 2+ Firecracker microVMs with no internet access. Works on x86 and even my M5 chip.
I was able to chat via 3 separate mechanisms:
① CPU contention
② Cache eviction timing
③ DRAM bandwidth contention
… across multiple CPUs, defeating all Firecracker's production recommendations (no SMT, no KSM, tenants pinned to different cores.)
Essentially, if shared hardware is involved, there's a good chance your agents can communicate covertly. And with really decent latency.
The good news: this is decidedly different from a *side-channel attack* where the agents could steal information (Meltdown/Spectre), a category of attack that providers like Vercel Sandbox mitigate. Further, in our real-world environment, these 'whispering' techniques are much harder to pull off and easier for us to monitor and detect (sampling hardware counters + lots of at-scale statistical data.)
Noteworthy: ironically, OpenAI and Anthropic's models refused to do this work. Open weight models got the job done. I'm worried the labs' stance might backfire and get in the way of helping developers build correct and secure systems. e.g: building a secure sandbox requires having this kind of analysis, and without it, we'll have a less secure world.
Is it just me or has LLMs responses has become very nerdy and terse bordering lab talk? Standard Technical English skill cleans that out but doesn't stick for long as usual.
Chinese developed tones for sound compression, and Japanese forced words into open vowels. Korean solved the puzzle with Hangul, a feat of phonetic design where shapes mirror the mouth and tongue, on top of highly expressive native spoken language.
Listening to same conversation improving incrementally by having my voice agent target audio from full-duplex voice agent demos as goal, basically:
speaker diarization -> transcript -> masked playback
Started with Tyger demos but plenty of 'teachers': GPT-Live-1, Dessa, etc.
