matt 🐢

the models keep changing. i keep building. notes from the trenches.

[matt 🐢 @m4ttbuilds]

the bar for paying for software just went way up. if agents can build a bespoke version of your app in under a day, you'd better be doing something they can't. i loved shottr (https://t.co/xlNfbHyO5z) but needed it off mac, so i vibed out sniplet, a cross-platform rust version: https://t.co/jdnYKUSUw8

[matt 🐢 @m4ttbuilds]

anthropic makes enterprises switch to api pricing instead of just giving every dev a $200 plan. that tells you how subsidized the $200 plan is

for a big org it'd save literal tens of millions

[matt 🐢 @m4ttbuilds]

@bemartinwong tying it to a reviewed commit makes sense. to the agent, a file someone slipped into the folder looks just as official as one you wrote

[matt 🐢 @m4ttbuilds]

a security firm found an AGENTS.md file in a customer's repo telling coding agents to install a rogue npm package. in their lab tests codex went along with it over 90% of the time

do you read a repo's AGENTS.md before you let your agent loose in it?

[matt 🐢 @m4ttbuilds]

openai shipped gpt-6.1 terra, a cheaper model for computer use that reads what's on screen and takes actions across apps

[OpenAI @OpenAI]

GPT-6.1 Terra: delivering fast and capable computer use at a lower price.

GPT-6.1 Terra understands what’s on screen, takes action across apps, writes naturally, and is a more capable model for your whole workday with the efficiency to take on more of it.

[matt 🐢 @m4ttbuilds]

@GetPingCron yeah that's the test i'd want too. if the agent can widen its own allowlist, it's not really an allowlist

[matt 🐢 @m4ttbuilds]

how do you let a coding agent use your api keys without it ever being able to read them?

a proxy that injects them on the way out? something else?

[matt 🐢 @m4ttbuilds]

@theo I thought you said something about Anthropic not liking that.

[matt 🐢 @m4ttbuilds]

tanstack just shipped redact, a smaller react runtime you add to your build config to ship less js without changing any app code. works with start, vite, vinext, and next.js

[TANSTACK @tan_stack]

Introducing TanStack Redact 🌴

A smaller runtime for your existing React app. Add it to your build config and ship less JavaScript without changing a single line of application code.

Works with TanStack Start, Vite, Vinext, and Next.js.

https://t.co/BzzTRpVTa1

[matt 🐢 @m4ttbuilds]

@maria_rcks @theo @t3dotcodes What did you doooo

[matt 🐢 @m4ttbuilds]

It’s inevitable! T3Code will take over the world. It’s the VSCode of the agent era.

[Matthew Berman @MatthewBerman]

I think I'm becoming T3-pilled. Having all my subs in one place, accessible from anywhere across multiple machines, and each agent being able to talk with the others is incredible.

[matt 🐢 @m4ttbuilds]

someone slipped a credential-stealing backdoor into a 7b model for under $50. it only fires on a secret trigger phrase

do you check where your local model weights came from before you run them?

[matt 🐢 @m4ttbuilds]

@theo still trying to convince my company that...

[matt 🐢 @m4ttbuilds]

deepseek v4.1 flash with the full official weights now runs on 4 dgx sparks, 1m context and 16 sessions at once

[Mia @MiaAI_lab]

Run DeepSeek V4.1 Flash 1017 with the full official weights on your 4x DGX Sparks, using TensorFold, Abliterated! 🔥

- Full support for switchless TP4
- 1M context, 16M total KV cache
- 16 concurrent sessions with 1M each
- 15 GB of free mem per Spark after load
- OFFICIAL weights, no further quantization!
- Needle passed at 100% depth
- ABLIT is off by default. Gated.

Performance tested on a switchless TP4. It should be faster with a switch.

Decode speeds, non-greedy, thinking on:
- 66 tok/s on prose, single stream
- 96 tok/s on prose, 2 concurrent streams
- 173 tok/s on prose, 16 concurrent streams

Decode speeds, greedy, thinking off:
- 85 tok/s on prose, single stream
- 113 tok/s on prose, 2 concurrent streams
- 210 tok/s on prose, 16 concurrent streams

Decode speeds per stream in deep context (128k), non-greedy, thinking on:
- 58.6 tok/s on prose, single stream
- 83 tok/s on prose, 2 concurrent streams
- 132 tok/s on prose, 8 concurrent streams

Prefill: 4000+ tok/s up to 256k.

This is the best model to run if you have 4x DGX Sparks, bar none! It works so well on the Sparks that it's unbelievable.

Get it here:
https://t.co/6Og2cWfW4f

[matt 🐢 @m4ttbuilds]

anthropic's new guidance says one main agent with skills and tools beats a tower of subagents handing work to each other

one agent or a swarm, which is working for you?

[matt 🐢 @m4ttbuilds]

@robinebers @theo @bot let me know if robin reports back on this topic anywhere

[matt 🐢 @m4ttbuilds]

google and anthropic both just put out free courses on harness engineering

is harness engineer a real job now, or prompt engineer with a new name?

[matt 🐢 @m4ttbuilds]

is your claude plan draining faster lately, or is it just the loudest corner of my timeline?

[matt 🐢 @m4ttbuilds]

@efbakir @theo They don’t each write code…

[matt 🐢 @m4ttbuilds]

what's the last paid app you replaced with something you vibecoded?

mine was shottr. built a clone in rust + gpui with an agent

[matt 🐢 @m4ttbuilds]

@joshmillgate How do you not have time? It’s vibe coding. 3 prompts and you got a working clone.

[matt 🐢 @m4ttbuilds]

@danielcberk Hope more people cancel so I can get mine sooner