Andrew

occupy mars!
slowly working on a WW2 game :

[Andrew @openmarmot]

@ParkSlopeFlngsm Odd comment considering the Obama library looks like the Berlin flak towers

[Andrew @openmarmot]

@ChuckCook Yeah I think that would be interesting. Would be cool to combine that with a more customizable cabin overheat mode

[Andrew @openmarmot]

@AndrewCurran_ I use grok every day to research software changes/github issues/software doc research. It is very good at real time data search. Might be SOTA in this niche. Hardly a failure.

Meanwhile LeCun only surfaces to let out more hot air. A very forgettable person.

[Andrew @openmarmot]

@SawyerMerritt His staff probably writes all this stuff. Doubt he is even aware of it

[Andrew @openmarmot]

@BasedTruthBomb @techdevnotes If you are using it in a harness the new grok build model is the way to go. I’ve been happy with it in opencode and especially with the new grok build cli

[Andrew @openmarmot]

@Jeyffre Not sure it would fit. Would have to be a q4 at least. With 4 you will need a loud network switch, and speed will probably be well under 20 t/s. Would be a fun experiment if you have the $$ though. I have a 2x spark cluster running minimax m2.7 at about 40 t/s. It’s pretty decent

[Andrew @openmarmot]

@pabis_eu @Jeyffre I have three and they are using 183 watts total at the moment. That’s basically nothing

[Andrew @openmarmot]

Some interesting FSD testing.
0.13 second reaction time.
Probably quite a bit faster as there is no way of knowing for sure at what point FSD identified the issue.

Current FSD is too reactive if anything imo. Mine keeps breaking hard for the elderly drivers around here who are constantly jump scaring it with their bad driving.

[ΛI DRIVR @AIDRIVR]

I calculated Tesla FSD's reaction time

[Andrew @openmarmot]

I've been working a bit more on writing a audio native ai harness.
speech to text -> LLM -> tool calls -> text to speech.

nothing serious, just a fun experiment.

the best thing about this is how simple the code is. the tool calling in particular works incredibly well and its just a couple lines of python, no silly overbuilt frameworks.

mostly built with grok build, which I have been enjoying quite a bit. it seems to have the multi-agent thinking that Grok 4.3 has, and the cli tool feels modern and fast compared to claude-code.

still aimed at hosting on a spark, which is a great platform for hosting light weight 24/7 cuda stuff like speech to text servers.

https://t.co/6ZkX0JsKL0

[Andrew @openmarmot]

@pmarca sounded interesting but not '$38 for the kindle version' interesting.

[Andrew @openmarmot]

@antirez hacker news comments have gotten really miserable and a lot of the submissions have become more political than technical in nature. i've stopped visiting it

[Andrew @openmarmot]

@SmashWallGames Grey was phased out for German vehicles very early in the war. If they need to be a single color they should be yellow-tan.

[Andrew @openmarmot]

@UnslothAI Does this mean that you fixed unsloth studio on the spark? Because that seemed kind of abandoned..

[Andrew @openmarmot]

@ryanvogel opencode should add a mic button/key and a option to configure a url for a whisper.cpp server. i wrote my own utility for this but i still have to paste in the resulting text

[Andrew @openmarmot]

@ryanvogel Kind of want

[Andrew @openmarmot]

@AndrewCurran_ meanwhile aws govcloud bedrock is still on sonnet 4.5. idk if its licensing terms or government approval that slows aws down, but it is vastly behind the commercial environments.

[Andrew @openmarmot]

@hagdaughter mcconnell's is the best that i've found so far.

[Andrew @openmarmot]

@vcoolish yeah i like grok build cli, but a web like interface like 'opencode web' allows so many more options for interacting and viewing stuff.

[Andrew @openmarmot]

@Jaaneek It’s good. Using it with opencode and grok cli with no issues.

[Andrew @openmarmot]

@Youssofal_ I disagree. Opencode is great. I use it on a regular basis with both local and commercial ai models with no issues