It would be fun to try and turn this into a RTL implementation and run it on a FPGA. The asset generation step may require a small CPU, but most other things including the renderer may translate well to a set of FSMs.
Super cool. I like how there’s little to no magic here. Just good ideas stacked on top of each other with sensible limits and careful choices.
I’ve decided not to travel international until my GC renewal is complete and I have a new card in hand. The risk, even if small is currently not worth it.
I’ve started thinking about local network storage and lightweight backups methods again for the first time in 15 years.
Small 10W PCs can drive a meaningful amount of network attached storage at acceptable data rates for really modest amounts of money.
Cool, you’re a lot further along than I am :) I did some work to build the drivers and original tests on Arm64 and cosim with the DosBox model. Happy to share if you’d like.
My late grandfather experienced what I imagine to be a similar paradigm shift in his career as a mechanical engineer, when CAD was introduced in the 1960s. I think about him frequently and need to write about this some time.
The other part is not wanting to depend on some cloud service to do critical parts of my work. I feel quite strongly about having the tools I depend on run on my local machine for as much as possible.
This is one half of my hesitation with LLM tools; I do a lot of my thinking while tinkering with the data model and types. It helps me find holes in my understanding of problems and how to apply the tools and libraries at my disposal. But me writing this is not the bottleneck.
Using a MCU would be a challenge due to a lack of fast SerDes. But an FPGA like Lattice’ ECP5 would certainly make SATA-2 at 3Gbps possible.
You could consider a dev board with a QSFP cage and design a QSFP receptacle with a DP connector on the other end. Depending on the dev board the adapter would be passive, but this assumes the QSFP side band signals are generic IOs and can be repurposed for AUX etc.
I am working on a board with similar motivation. But it’s still many months away from even a prototype.
My hope is that model requirements for specific use cases will trend down and that unified memory bandwidth in SoCs from Apple, AMD and Qualcomm will continue to increase.
I don’t know if for example we’ve yet tried to optimize a model architecture for single user throughput/latency.
I am not saying there isn’t anything in the roadmaps, but maybe with the exception of Apple and Qualcomm everyone else is still very much focused on large data center applications. Client side LLMs seem to be a byproduct, not a focus.
Unlike more traditional ML models (ranking, image recognition, etc., which these NPUs were targeting), LLMs are less about TOPS and more about memory bandwidth, especially in sparse model architectures. Compute often sits idle, waiting for data.
I’ve been using their assistent in research mode occasionally since it was launched. The integration with their search engine seems to work quite well and my experience has been surprisingly good/useful.
It’s not just Europe not buying US weapons anymore, but those EU contractors will start competing for the same remaining market.
If these companies figure out how to scale and compete (not a given unfortunately) they may very well run away with a chunk of additional LATAM and EMEA business.
Commute is a real drain on both productivity and happiness. On days I am in the office I spend 3 hours in the car between commute and kids school drop off.
I’ve been told not to complain, prior generations managed. But with young kids, both parents working there’s not much gas left in the tank.
www.mindshare.com/ebook/pci-ex... is a good i intro. There may or may not be some PDF copies floating around on GH.
I wasn’t and still may not be. But this @acquiredfm.bsky.social interview with him is excellent and he certainly earned some respect for his willingness to reflect on his actions in the open.
I should have hoarded RAM instead of investing in the stock market.
This is a depressing but excellent presentation for anyone who has (young) kids and teens.
The internet isn’t the place anymore we grew up with and navigating this with our kids is mandatory. And unfortunately generative AI will only super charge this further.
That’s very much my intend. Starting early means we won’t have to cover it all at once and my hope is that by participating and providing her with high quality options we’ll have a trust base from which to start building the more serious stuff.
Multiplayer currently means us on the couch playing games together.
Most of the conversation has focused on how to spend time and attention, the patterns in free games to get and keep you hooked, why everything cost money. It’s a start.
Fortunately we haven’t gotten to the multiplayer stage yet, but yes it needs to be part of the conversation. I love the line from the presentation that “friends are people we know in real life.” I definitely plan to use this.
I’ve started having conversations with my 8yo about free games, the motives and patterns behind them and why I encourage her to play alternatives.
Its work in progress, but even at her age they understand a lot. As a bonus we’ve been playing more games together. Providing higher quality fun is key.
This talk is excellent. Depressing, but excellent. And the presenter is right; the only way to protect kids is to develop a relationship with them which is stronger than what they can get out of the apps. That takes time and effort. But unlike maybe our parents we can’t afford not to.
Original Pokemon carts for these systems are getting pricey, but original GBA carts can still be had for ok-ish prices in case you want to get him some. The GBA has excellent ports of NES, SNES and some Sega classics.
+1 It’s not cheap, but a good experience out of the box. There is plenty of opportunity to tinker and get more out of the system later, but that’s not at all required in order to enjoy the system 5 minutes after unwrapping it.
What would be interesting to do in order to improve latency/accuracy of a software emulator is to build something more or less directly on top of the Linux kernel and explicitly boot that as a dedicated environment. This should greatly reduce timing noise from running background tasks etc.
