This isn't your typical acquisition blog post. I'm really excited not just that @rough__sea, @piscisaureus, and friends are joining, but also what they plan to work on: Making workerd a first-class, self-hostable open source runtime. Read for more:
https://t.co/VfBtRyg6k6
@rakyll Wouldn’t be an issue except they only land them publicly in batches on releases rather than merge into public main (not to mention the verification suite being private)
@thdxr are you finding that codemode is more token efficient than lazy tool loading/tool search alone? i know theoretically it should win for chaining tool calls, but realistically i'm seeing it as a wash when you consider the model iterating on the code
@NathanFlurry Have you actually found code mode to be more token efficient than just lazy tool loading alone? I haven’t been able to replicate those results. I understand the theoretical benefit when tool chaining but the overhead of the model writing code and iterating has been dominant.
@rough__sea As a fan of the project, early adopter, and bug finder, I would love if you ported commits upstream between releases and shared more openly the internal dev machinery, even if you want to keep contributions restricted. Building on it is hard when changes come all at once
@NathanFlurry Do you mean they share isolates for a single DO across requests? Presumably it is not secure to share across different DOs though that is what celld is doing right now (using realms). How are Rivet Actors doing secure isolation between tenants?
@jhleath Yeah I think you did spend too long at AWS 😄. There are so many counter examples of this. Docker, Mongo, data bricks, hashicorp. Not to mention the open core companies. Or Linux distributors. It’s been a legit strategy. Not sure what the future holds.
Finally you can extend Envoy with native Rust (but not limited to it) shared libraries! The work has been supported by multiple companies including Netflix, Linkedin, etc. Looking forward to more users and use cases! https://t.co/fr6p35Da8E
At Cruise I built robotaxis that completed 250k driverless rides, including the first rides ever in a major city.
With the Tesla event coming up, many people have asked for my thoughts. I truly hope they get robotaxis working. Autonomy is the fastest and most effective way to eliminate car accidents, and we desperately need this.
I don’t know what we will see tonight. What I know is that it takes a non-trivial amount of work to go from making a car mostly drive without interventions to safe, robust, and legally compliant robotaxi network that meshes well with local communities.
So here are 15 key things to look for from new robotaxi players:
1) Getting stuck - is there a method for remote operators to relocate unoccupied vehicles that are blocking traffic or emergency vehicles? What happens if there are no available remote operators and the vehicle becomes stuck?
2) Detecting collisions - is there a high recall system that can detect collisions, including minor contact with cycles or pedestrians? Does this comply with local, state, and federal reporting requirements?
3) AI override - is there a way for remote confirmation of critical, long tail decisions (is it safe to stop on these railroad tracks, was that a collision or not, is it ok to enter this emergency scene, is that officer telling me to stop)?
4) First responders - is there a way to remotely unlock cars and provide access to first responders who need to relocate a stuck vehicle? Is there a training program in place? Is there live phone support?
5) Connectivity dropouts - do these cars have Starlink, cellularly redundancy, or some other way to remain connected? What happens if telemetry is lost for an active but unoccupied vehicle and it gets stuck?
6) Sensor cleaning - is there a way to clear blocked or dirty sensors? Are various forms of degraded performance detected and mitigated?
7) Degraded states - in the event of a computer, sensor, or software failure, can the vehicle utilize a backup system to safely pull out of traffic or otherwise reach a safe stopping location? Is there coverage for all known or plausible faults? Will this work properly on highways where pulling over is not necessarily a safe option?
8) Congestion control - will vehicle take diverse routes to avoid creating traffic jams? If 20 vehicles arrive at a busy concert venue is there any “air traffic control” to ensure they don’t get stuck in a cluster?
9) Emergency vehicle detection - do vehicles correctly pull over or yield to emergency vehicles? Can they traverse active emergency scenes when necessary or appropriate?
10) Long tail detections - do they avoid flooded areas, downed power lines, wet cement, caution tape, crossing guard hand motions, open pits or manhole covers?
11) Liability - who is at fault when a vehicle causes property damage or injury? Are there high recall data logging systems in place as needed to absolve the owner of liability when the other driver was truly at fault?
12) Regulation and permitting - will these vehicles operate in states with public reporting requirements and meet all requirements? If vehicles do not have a steering wheel or traditional controls, will they be self-certified by Tesla as FMVSS compliant?
13) Bad weather - does the system correctly degrade its performance in the event of sudden changes in weather? What happens if a trip is in progress when weather becomes severe? Will it refuse to operate if conditions are too severe?
14) Pullovers - does the system avoid pulling over in bus stops, restricted areas, or in front of private driveways? What happens someone needs the car to be moved from their driveway?
15) Local laws - does the system obey local traffic laws, and who pays the ticket if there is a violation?
Tesla will undoubtedly solve all of these eventually. Have they done it yet? Who knows. I’m still very excited to see what they’ve been cooking up.