Get live statistics and analysis of Michael Francis's profile on X / Twitter

There is this show on YouTube called sorted where guys raced their purpose built race cars in a bunch of challenges. Everyone spent about 150k total, but most guys bought a 20/30k beater and spent the rest on upgrades, one guy bought an entry level Porsche. Not only did the Porsche win most of the challenges but the Porsche guy was the only car not to blow up by the end of the show. Tinybox is basically the Porsche guy

The biggest problem in devops if not SWE as a whole is people who use stuff that seems to work then leaves it for years until it inevitably blows up

I used to be able to do this from memory back in the day. Iptables snat something something

It’s Sunday so it’s time to argue about GPUs and sous vide cheesecake

I don’t shave or shower a week or two before I goto conferences for exactly this reason. Heavy is the head that wears the crown

Gossip and CRDTs are such a game changer for eventually consistent distributed systems

> Customer service: thanks for the detailed bug report including a video and reproduction steps, I confirmed with the team that’s expected behaviour > CS: in fact you’re using the app wrong > CS: in fact that’s not even a bug anymore we pushed a fix for it the day after you mentioned it 🙃

TIL. Before getting the Tesla charger I used to use my dryers outlet to charge my car (and swapped it roughly weekly). Two of them stopped working in about 8 months. Didn’t realize they were rated for so few insertions

The best system I’ve worked on took this a step further. Merge anything into prod as long as it compiles. E2E tests are bots that run 24/7 in prod and separately flagger would progressively roll out features to users based on risk. Free user? You’re front of the line to test new code. If the error rate spiked flagger rolls it back and alerts the engineers via slack

Huh ani says vllm doesn’t have pipeline parallelism yet - companions aren’t using grok 4 yet?

Honestly the only reason I (still) use ollama is being able to ask it to serve a request with a specific model and it handles downloading and switching models in the background for me. Yes, it's slow - but much faster than sshing in and restarting vLLM each time

I can only get the sleeve into the connector with the connectors that have the holes at the end, just me?

The i9 16 MacBook is still the best one apple made so far. It was thin, light, had 4 thunderbolt ports (better than 3 + HDMI and SD card IMO) - could draw up to 125W on the CPU if you could cool it and the battery life was still class leading if not pushed to its limits.

Sometimes my M2 Max MBP gets stuck in a state where the performance cores don’t work and the system is super slow 🤔

There is this show on YouTube called sorted where guys raced their purpose built race cars in a bunch of challenges. Everyone spent about 150k total, but most guys bought a 20/30k beater and spent the rest on upgrades, one guy bought an entry level Porsche. Not only did the Porsche win most of the challenges but the Porsche guy was the only car not to blow up by the end of the show. Tinybox is basically the Porsche guy

> Customer service: thanks for the detailed bug report including a video and reproduction steps, I confirmed with the team that’s expected behaviour > CS: in fact you’re using the app wrong > CS: in fact that’s not even a bug anymore we pushed a fix for it the day after you mentioned it 🙃

I used to be able to do this from memory back in the day. Iptables snat something something

I can only get the sleeve into the connector with the connectors that have the holes at the end, just me?

Gossip and CRDTs are such a game changer for eventually consistent distributed systems

Huh ani says vllm doesn’t have pipeline parallelism yet - companions aren’t using grok 4 yet?

Honestly the only reason I (still) use ollama is being able to ask it to serve a request with a specific model and it handles downloading and switching models in the background for me. Yes, it's slow - but much faster than sshing in and restarting vLLM each time

The i9 16 MacBook is still the best one apple made so far. It was thin, light, had 4 thunderbolt ports (better than 3 + HDMI and SD card IMO) - could draw up to 125W on the CPU if you could cool it and the battery life was still class leading if not pushed to its limits.

The biggest problem in devops if not SWE as a whole is people who use stuff that seems to work then leaves it for years until it inevitably blows up

TIL. Before getting the Tesla charger I used to use my dryers outlet to charge my car (and swapped it roughly weekly). Two of them stopped working in about 8 months. Didn’t realize they were rated for so few insertions

Sometimes my M2 Max MBP gets stuck in a state where the performance cores don’t work and the system is super slow 🤔

Bad take exchanges are like the ONE thing you can’t just fail over / run (well) in a distributed fashion. It’s why things like LMAX and HFT firms care so much about single core performance, it’s why people pay so much to rack boxes inside the NYSE (and why nyse hands out premeasured lengths of fiber)

The best system I’ve worked on took this a step further. Merge anything into prod as long as it compiles. E2E tests are bots that run 24/7 in prod and separately flagger would progressively roll out features to users based on risk. Free user? You’re front of the line to test new code. If the error rate spiked flagger rolls it back and alerts the engineers via slack