James Fang

@mehran__jalali My team has spent a lot of money too running evals and just getting eval infra to work correctly. But it’s necessary work to validate your training and / or agent design went correctly!

I never realized just how much running benchmarks costs. Evaluating Opus 4.7 on MMMLU *once* costs $25k-$50k

It’s not sexy, but this is where good software engineering practices have extreme utility in ML systems and infrastructure.

@tengyanAI a16z is the last investment firm that I’d expect to invest in Anthropic, their philosophies are diametrically opposite

a16z not being in Anthropic is one of the biggest omitted investments in AI

One (probably controversial) observation I’ve seen in B2C is that unfortunately there I can’t think of many products that are successful, that don’t have an offering to an enterprise version (perhaps offered by a competitor but the fundamental core product is the same), without appealing to a human vice or a human psychological flaw, such as: - lust (anime chatbot apps) - greed (gambling apps) - poor attention management (social media)

@anothercohen Mattermost is the move!

I'm tempted to finally churn off Slack. We're paying ~$6k/year for 40 people and they just quoted me $21k/year for the business version that includes a BAA (and all the shitty AI features). Incredibly overrated software

As a ML infra / systems engineer, from my experience and chatting with other engineers, the easiest way to make every engineer get extremely angry is to introduce rate limits without warning. Yes, good rate limiting policies will take effort to develop, and requires disciplined resource management. But it is imperative for developer trust. And engineers hate nontransparent hidden rate limits a lot more than even sudden pricing changes, because instead of a surprise bill engineers would be waking up at 2 am to debug for hours why their systems just broke.

You have to understand. Spending 1 hour per day on brainrot is insane. That's about 6% of your waking day. About 5 years of your waking life. Half a decade. On brainrot. Just gone. Zero return. Zero fulfillment. Zero meaning. Zero contribution to the other parts of your life.

Most people don't get burned out from real work. They get burned out from fake work. The constant performance of productivity without tangible output.

I’ve been in New York for 5 days, what stands out to me is that New York seems significantly less divided and less hostile to young people than SF. An example is that plenty of housing is getting built where young people live in Western Queens and Northern Brooklyn.

@handzelmatt Would be even more impressive if it wasn’t Claude but instead Claude Code lmao

@handzelmatt Instead of Dr Google we got Dr Claude

Claude correctly diagnosed me with low iron I've been writing more in detail about my sleep and how I feel on runs Over the course of a week it analyzed these entries and suggested I get a ferritin test Got my test results back today and I'm deficient

Will be in New York on Tuesday morning the 7th to Monday Evening the 13th. What places should I check out? Btw, I am a MLsys/Infra engineer, and I love: - Classical Piano - Chess - Biking - Hiking - Gym If there’s a social / dating scene I should check out too, please drop it in!

@jimkxa @tenstorrent Truly not enough people understand how much interconnects bottlenecks are a pain in the ass, you only understand it when dealing with massive training (maybe not even inference) deployments

I’ve been thinking about the usual order of problem solving in computers 1) Compute 2) Memory 3) IO Scaling with MPI or similar is often an exercise for the reader @tenstorrent solved it in this order 1) Data placement and movement for Scaling 2) IO 3) Memory 4) Compute Compute is the easiest and best understood so making it last makes sense This was hard, actually, but puts the hard part first. Results are surprisingly good

Need a place to live after my lease is done in July in Palo Alto. Got any recs? - prefer somewhere in South Bay - happy to do roommates - ideal budget <$2000 a month, hopefully <$2200 a month

bit late to the recruiting cycle, but looking for a summer internship in ML/hardware/inference!! i've been working on CUDA kernel writing, FPGA acceleration and RTL. would love to find a team doing similar work this summer dual US/Canada citizen, can relocate anywhere DMs open :)

@krupaad @LaurenceLiang1 Saw this post from one of my mutual’s repost’s, gonna repost it too to support you too. Good luck!

Not to be the downer, but articles like these are inviting people to find and exploit their security vulnerabilities, I hope someone helps them with the security side

@GoobertownA I’m in Palo Alto wya

@polymath_james Me

Who felt the earthquake just now