John Lam

I work on building our AI coding experiences @Microsoft. Accidentally co-created GitHub spec-kit a few months ago (

[John Lam @john_lam]

2026 will be the year of tools for agentic thought

[John Lam @john_lam]

@ashleevance Sounds like an amazing father and a life well lived. Thank you for sharing this and I hope you find comfort in his memory.

[John Lam @john_lam]

@brupm @GergelyOrosz i came here to say the same thing. love mine.

[John Lam @john_lam]

@ENNEPS @r3a9an_k_ Acceleration boost unlocked!

[John Lam @john_lam]

how do you acquire the data though. that is a giant pita. i used tiller but that is $80/year spend - spend that i already use on my chatgpt pro sub so happy to not add incremental spend. plaid directly was a weird pricing model. no way i want to maintain the data ingestion pipeline myself- life is too short for that.

[John Lam @john_lam]

@petergyang @ChatGPTapp i did this by adding a manual account. after ingestion it asks you if you want to add these accounts. from that you can generate dashboards to track goals etc.

[John Lam @john_lam]

plenty of "sane" people use quicken, plaid, tiller and other personal finance services/apps. this is read-only access, not transaction level access. the attached blog post says they use plaid today with intuit coming later. i vibe coded a personal finance app using tiller to ingest data, store it locally in sqlite, and then i run agents over it. this might replace that for my use case depending on how good it is (i like my custom dashboards).

[John Lam @john_lam]

@steipete @buildwithhassan context window of BRAIN is 7 +/- 2 right?

[John Lam @john_lam]

@badlogicgames @FrameworkPuter @antirez is this the max+395 w/128gb? pls post pics once you have it assembled!

[John Lam @john_lam]

@PatrickMoorhead @Picomoorhead @nvidia congratulations!

[John Lam @john_lam]

@PatrickMoorhead do you know whether this is 18A or 14A? seems more likely to be 18A if this is a near-term first volume production test?

[John Lam @john_lam]

@SamPeterToT @KobeissiLetter this is not about x86

[John Lam @john_lam]

be careful about model being trapped in a local minima. it is quite expensive and time consuming if it does. in my case it reinforced itself to breaking large files into single class/entity files. lots of activity for very little benefit. figuring out how to get it to reflect better on "so ... were the last n commits helpful or harmful? steelman both sides and figure out if you need to course correct"

[John Lam @john_lam]

another, harder, problem is reviewing the implementation. reading the code does not scale to human level inference speeds, ie you are bottleneck. i wonder if we can get agents to write "e2e manual verification programs" such that output is easy for human to validate (in addition to tests etc of course)

[John Lam @john_lam]

i don't think waterfall can be run in minutes because humans cannot review the giant specs in minutes. i'm beginning to think that "progressive, iterative prototyping" might be an interesting path to explore because reviewing behavior is faster than reviewing a spec about behavior.

[John Lam @john_lam]

@stuartbuck1 one of my high agency interns cold emailed steve ballmer and had lunch with him years ago. you would be surprised by how little direct to him emails that he gets. something like 50-60 a day.

[John Lam @john_lam]

@davidfowl the entropy in your codebase compounds over time. humans still needed for this task (for now).

[Nathan Lambert @natolambert]

Opus 4.7 has a new tokenizer.
This means it's also a new base model.
Glory days of pretraining still very much going.

[John Lam @john_lam]

@bernhardsson @_gongy That's a nice 991.1! NA for the win.

[John Lam @john_lam]

@RhysSullivan i really like this slide from @mitsuhiko. writing small self contained libraries and composing them might be an interesting avenue to explore. deck is https://t.co/jwnWHAbS91

[John Lam @john_lam]

@nicopreme this is a cool idea. but i wonder about the generality of the approach - have you thought about what it would take to build something like this but in code? i would imagine that a pi extension comes really close to being able to do this.