no passengers on the battleship. founding CEO @Mapbox.
Suisse
Joined March 2007
- Tweets14.6K
- Following1.9K
- Followers10.4K
- Likes23.8K
Pinned Tweet
the first @desertantlabs models are now live (12 stable and six in beta), via one SDK for Swift, Kotlin, + JavaScript. one model per task, each built to be the fastest way to complete that task on a device: desertant.com/models/
but the microwaves..... “AI worms” will soon be able to hijack smart home devices such as ovens and TVs, creating a new cyber security threat that will be difficult to contain. The developer of Claude said in a blog post on Monday that AI agents may be months away from “self-replicating” across everyday household devices, to create potentially uncontrollable networks of AI agents. giftarticle.ft.com/giftartic…
installing @desertantlabs CLI
+ speed tests transcribing videos w/ Voz(and batch transcription). then bending Vox vs whisper.cpp
@joemaddalone shows workflow: extracting shorts from videos w/ Clips ➡️ audio cleaning w/ Clear ➡️ identifying fillers w/ Uhm ➡️ generating titles
Tiny models are dope.
youtu.be/-89RZXTvY1k
the team @desertantlabs is growing: desertant.com/jobs/
+ ML Researcher
+ Cloud Infrastructure Engineer
+ Runtime and Performance Engineer
+ Mobile Engineer: Android
+ Developer Relations Engineer
+ Partnerships Manager, Europe
... 🐜
we’re in orbit! @PAVESpace 1st mission of the in-house avionics architecture to space just reached LEO. Next the team will start self-health checks/in-orbit reprograming, and then run the navigation algorithm assessments withstand high-radiation environment of the Van Allen belts that the next heavy-lift kick-stage will fly. Awesome work Julie Böhning, Jérémy Marciacq and the entire team pave.space 🇨🇭🚀🪩
has anyone played w/ new cli to if it's easy for @grok pull data updates updates? like @USGS naming changes from @StateDept --> directly into studio @mapbox
the design work by @jgebbia is 💪
eric gundersen retweeted
Jev + on-device models = results in seconds with no LLM in the loop.
Quick demo app to show the possibilities. Drop in an audio file: Ear detects the language, Voz transcribes it and Redact removes PII.
Then @typesafeai's Jev makes about 20 decisions in one call in milliseconds, and picks which of our on-device models to run. Voice memo to to-do list. Meeting to redacted transcript. Podcast to clips.
more granularity @Cloudflare "Block AI Bots" is being deprecated in favor of separate Search, Training, and Agent controls set at the domain level readwise.io/reader/shared/01…
eric gundersen retweeted
Do agents call 911 when given the option?
github.com/yhahn/oai-hf-rese…
Research on model escalation paths in the OpenAI-Hugging Face Incident
METRs report noted that agents didn't escalate to humans in the OAI-HF incident, but didn't dig into if and what pathways they had to do so. I take a look at how agents perform when given different pathways to escalate, and what we can learn from human safety systems like 911 that make escalation stupidly simple.
“big complex open-weight models become pretty indistinguishable from big complex closed-weight models, can be accessed at a fraction of the cost, and come with riders that are valuable in their own right” via @FT stats re open-weight and closed-weight AI models by release date and artificial analysis intelligence index ft.com/content/48588acb-8026…
OpenRouter API stats; four week total tokens used on open-weight or closed-weight models. ft.com/content/48588acb-8026…
Dragi is going to love local on-device AI ft.com/content/f054f927-b512… re FT op-ed today:
re EU sovereignty stats from Dragi’s FT's op-ed:
“The EU hosts under 5 per cent of the world’s AI compute versus 75 per cent for the US. Even for ordinary data centre capacity, the gap between demand and installed supply in Europe is estimated at around 3GW, roughly a quarter of what Europe currently has, and is expected to widen to 14GW by 2030.”
“More expensive electricity makes building and powering a facility in parts of Europe twice as costly as in China.”
ft.com/content/f054f927-b512…
eric gundersen retweeted
The new DeepSeek V4.1 Flash model is mindblowing - back on top of the open-source model leaderboard and extremely cheap.
It has a lot of very smart ways to be efficient and highly capable so I made a video of the forward pass to give you a view of what going on inside the model during inference.
Read more at huggingface.co/deepseek-ai/D…
And find the weights at huggingface.co/deepseek-ai/D…
“We had to turn to an open-weight Chinese model to process the attack logs. Many people have suggested that open-weight AI models are a threat — that they will be used to attack systems protected by closed-source AI models. That weekend, the opposite happened.”
great FT op-ep by @Thom_Wolf
ft.com/content/9faf688d-9192… What we learnt from OpenAI’s hack of Hugging Face
featured in Apple's keynote today @desertantlabs "Detail uses the new core AI framework and runs its AI models on device to make content creation fast and easy on the big screen."
