Mind n Machine 🦋

Building apps & workflows with Claude Code
Side Effects: Surreal Storytelling and Cool content

[Mind n Machine 🦋 @mindnmachinex]

https://t.co/Sg3N1bsgYs

[Mind n Machine 🦋 @mindnmachinex]

. @AnthropicAI is a company everyone needs to be watching AB’s learning from ⚙️

[Claude @claudeai]

We're introducing Claude for Teachers: free access to premium Claude capabilities for verified K-12 educators in the US, with a library of teaching skills and a direct connection to evidence-based curricula, mapped to academic standards in all 50 states.

https://t.co/5hZZijVPCV

[Mind n Machine 🦋 @mindnmachinex]

the model my business runs on just started building its own chips and filed to go public

do you trust a tool more once it starts building its own hardware, or does that make you brace for what's next

anthropic's in talks with samsung on a claude-tuned chip, sitting next to the S-1 it filed after a $65b raise. every other lab already has a silicon program, claude just caught up to itself

[Mind n Machine 🦋 @mindnmachinex]

apple says openai's hardware chief told job candidates to bring 'actual parts' to interviews

trade secret theft, or just how hiring wars actually work now

filed july 10th: a second ex-apple employee allegedly downloaded dozens of confidential files while building openai's hardware line, the one built on jony ive's $6.4b studio

the model war gets the headlines. this is what the talent war actually looks like up close

[Mind n Machine 🦋 @mindnmachinex]

got a "partnership" email this week. free pro access to an AI music tool plus 30% commission, sounds fine right?

i had my claude pull the thread, search the tool, search the agency, then check what other creators were saying about the exact same email

five minutes in: real-ish agency, real tool, zero actual pay, and a warning from another musician who already got burned by this exact email

declined. draft written, nothing sent, under ten minutes total

[Mind n Machine 🦋 @mindnmachinex]

would you leave the team that made you famous for the team a rival built?

the guy behind alphafold just did. john jumper walked from google deepmind to anthropic after 9 years, and two more gemini researchers followed him the same week

talent doesn't leave for a paycheck at this level. it leaves for who's shipping

[Mind n Machine 🦋 @mindnmachinex]

what does an mit license on a 744b-param model actually buy you?

glm-5.2 from z ai is on hugging face now. no commercial restrictions, no regional limits, no acceptable use policy. fork it, fine-tune it, embed it in a product without asking permission.

→ 1m token context, built for repo-scale coding and agentic work
→ terminal-bench 2.1: 81.0
→ swe-bench pro: 62.1

the constraint hiding inside that openness: cluster-scale h100 compute to run the full model. self-hosting is off the table for most teams.

for labs and hyperscalers, the mit license is the actual unlock. for everyone else, the play is the api, not the weights.

know which scenario you're in.

[Mind n Machine 🦋 @mindnmachinex]

the workflow most people skip on AI image tools:

write a character file
exact features, tattoos, outfit, lighting, posture, room

paste the entire description into every single generation

the consistency isn't a setting you toggle on
it's a document you re-feed

skip this and your character is a different person
across cover, thumbnail, music video, social posts
the audience can tell

[Mind n Machine 🦋 @mindnmachinex]

https://t.co/4pf0RkCpiq

[Mind n Machine 🦋 @mindnmachinex]

prompt for killing a plan's blind spot before you ship it:

paste your plan, then:
"argue the most likely way this fails that i'm not seeing. make the case like you want me to cancel it."

→ the model isn't attached to your plan, so it'll say the thing your team won't
→ most plans die of the one risk nobody wanted to name

bookmark to use it later ⚙️

[Mind n Machine 🦋 @mindnmachinex]

how context windows actually fail isn't size, it's position

models read the start and end of a long context well and skim the middle

→ bury the key instruction at line 4,000 and it quietly disappears
→ "lost in the middle" is measurable, not a vibe
→ a 1M window doesn't fix it, it just makes more middle

put the load-bearing context first or last. never in the soft center.

[Mind n Machine 🦋 @mindnmachinex]

claude code shipped artifacts (beta, team + enterprise): your coding session publishes a live html page at a private org url that auto-refreshes while the agent works.

→ pr walkthroughs with annotated diffs
→ live dashboards from session data
→ investigation timelines, checklists, bug walkthroughs

the session and its documentation become the same artifact. not a log, not a screenshot. a browser page that updates in real time.

csp blocks external calls. org-scoped access only. claude asks permission before publishing.

execution traces become shareable ui. that's the pattern.

[Mind n Machine 🦋 @mindnmachinex]

you can't benchmark a model you never configured. most people are doing exactly that.

[Mind n Machine 🦋 @mindnmachinex]

microsoft built its own reasoning model from scratch, no distillation

mai-thinking-1, announced at build

→ sparse moe, ~35b active params, 256k context
→ trained on licensed/enterprise data, not distilled from another lab's model
→ one of 7 in-house mai models: code, image, voice, transcribe
→ the point is a first-party stack, less dependence on one provider

building your own frontier model is now a platform decision, not just a research one

⚙️

[Mind n Machine 🦋 @mindnmachinex]

gemini cli dies today for free, pro, and ultra users. antigravity cli is the replacement.

what changed:
→ go-based tui (was node.js)
→ multi-agent and async tasks are first-class
→ hooks and extensions are now "plugins"
→ closed source; gemini cli was apache 2.0

if scripts or ci/cd call the gemini command on those tiers, they break now.

this isn't a version bump. google is running a platform play.

[Mind n Machine 🦋 @mindnmachinex]

the trick that makes an AI vocal sound produced instead of generated:

don't use one take

→ generate the lead vocal
→ generate a second pass of the same line, prompted slightly breathier
→ pan them a few degrees apart, drop the second one low in the mix

the ear reads the tiny mismatch as a real double-tracked human
one clean take always sounds like one clean machine

the realism is in the imperfection you add back

[Mind n Machine 🦋 @mindnmachinex]

what's the one task you won't hand to AI, even though you could?

that thing is probably the work you actually care about

watch what you delegate without a second thought
and what you guard

→ the guarded thing is your real craft
→ everything else was just labor you'd tolerated

it's a mirror for your priorities
it shows you what you were willing to give away
the moment giving it away got easy

[Mind n Machine 🦋 @mindnmachinex]

the habit that catches the most AI mistakes:

never accept a large change you didn't read in chunks

→ the model writes 200 lines confidently
→ 195 are fine
→ 5 quietly change behavior you didn't ask about

read the diff in passes, not as a wall

the bug is never in the lines you'd expect
it's in the one that looked like cleanup

[Mind n Machine 🦋 @mindnmachinex]

Anthropic just quietly circulated a policy memo about what happens when AI hits exponential acceleration.

and the most interesting part isn't the technical roadmap.

it's the section on coordinated pauses.

the memo talks about conditions where labs would need to voluntarily slow down, not because the government told them to, but because the risk calculus inside the lab changed.

it also mentions labor displacement as a first-class concern. not "jobs will change." not "upskilling." specifically: what happens when the pace of displacement outstrips the pace of adaptation.

this is the company that's building the thing saying: we might need to pump the brakes, and here's what that would look like.

what's your read: is this a real policy position or is it optics?

[Mind n Machine 🦋 @mindnmachinex]

when your agent is a directory, it becomes a git repo.

vercel shipped eve today, open source. the framework maps files to capabilities:

→ agent/instructions.md → system prompt
→ agent/tools/*.ts → one file per tool, filename = tool name
→ agent/subagents/ → nested delegation
→ agent/channels/ → slack, discord, http, cron, same agent

handles durable execution, sandboxed compute, and human-in-the-loop pauses out of the box.

like next.js conventions, applied to agents. define the capability surface by dropping files. the framework wires the rest.

diff it, fork it, version-control it.

npx eve@latest init my-agent

[Mind n Machine 🦋 @mindnmachinex]

if you never object, people stop expecting you to
if your needs never surface, people stop asking
if you're always fine, people stop checking

the fawn response works so well
it makes you disappear

and the people who love you help
because you taught them how

[Mind n Machine 🦋 @mindnmachinex]

https://t.co/2lyKUJ9Vvf

[Mind n Machine 🦋 @mindnmachinex]

prompt for turning a vague doc into a checklist you can't fake:

paste the doc, then ask claude:

"convert this into a checklist where each item is either done or not done, no maybes. if an item could be marked done while the work is still incomplete, rewrite it until it can't."

→ vague checklists let you lie to yourself
→ binary items don't

bookmark to use it later ⚙️