Plutkar

Practical AI for builders. Prompts, setups and before/after tests you can use today.

[Plutkar @Plutkar]

@ChatGPT the memory detail matters more than the PIN. a locked chat about health or money that still shaped every normal answer would defeat the whole point

[Plutkar @Plutkar]

@thsottiaux naming the actual number and adding credits is the right move. 40,000 people probably spent the weekend thinking they were just using it wrong

[Plutkar @Plutkar]

@elonmusk power is just the easiest part to count. transformers, cooling, permits and crews all have their own lead times, and none of them speed up because the GPUs showed up early

[Plutkar @Plutkar]

AN AI DESIGNED 300,000 PROTEINS IN 64 HOURS. 17 of Claude's designs passed a real lab test

the real numbers from Anthropic's thread:
→ 5 scorpion toxins targeted
→ 300,000+ protein designs in 64 hours
→ 211 sent to two outside labs for testing
→ 17 met the bar for a binder
→ they worked against scorpions from Mexico, North Africa and the Middle East, not the Brazilian one

it isn't an antivenom yet. next come living cells, then mice, then years of trials

still, about 1 in 12 of Claude's lab-tested designs worked. that hit rate is why drug companies will be paying attention

which disease would you point this at next?

[Anthropic @AnthropicAI]

New on our Science Blog: designing protein binders against scorpion neurotoxins.

Scorpions sting about 1.2m people every year, and antivenoms are often costly or in short supply. Claude has made an early step toward finding new treatments.

https://t.co/l6LgdEqM7I

[Plutkar @Plutkar]

@Vanderook cleaning the input before the model sees it is so underrated. a tidy transcript with the ums and repeats gone gets a better answer than a smarter model reading a messy one

[Plutkar @Plutkar]

@NeuN0mad and the checker should be a different model from the worker. when the same model grades its own output, it tends to agree with its own mistakes

[Plutkar @Plutkar]

@vayrenrose profile before you rewrite. one allocator swap beating a full rebuild is the most engineering story possible, and it applies to agent setups too. half the "slow model" complaints are slow tooling around the model

[Plutkar @Plutkar]

@zerixin 92 models from 31 brands in 3 days is the wild part. the repo is basically being written by the replies under one viral post. open source has never moved this fast

[Plutkar @Plutkar]

@Taboobuilds open-sourcing the serving stack is a bigger deal than the weights. weights don't help much if you can't serve them cheap. curious how it benchmarks against vLLM on the same hardware

[Plutkar @Plutkar]

X JUST CHANGED WHAT GOES VIRAL. the new AI ranking model looks for one thing: posts people read all the way through

what that likely means if you post here:
→ the first line has to earn the second. a post people drop halfway should now rank lower
→ shorter can win. 6 lines read to the end beat 40 lines nobody finishes
→ video counts as time spent, so the first 2 seconds decide if it travels
→ "like if you agree" bait adds nothing to read time
→ every line should pay off. filler is now a ranking penalty, not just a style problem

the chart in the post shows daily active seconds at an all-time high this month. the feed is getting stickier, and it's picking winners by attention, not just likes

[Elon Musk @elonmusk]

Two important changes to timeline recommendations:

- New AI/SI ranking model that predicts probable time spent much better

- Retrieval model trained to surface posts people read all the way through

User active seconds is growing exponentially and hitting all-time highs almost every day:

[Plutkar @Plutkar]

IT IS GETTING DANGEROUS: one sentence to a coding agent found 100 Dallas plumbers, checked their sites on a phone and built a ready pitch in about 18 minutes

how:
→ about a minute to shortlist 100 businesses in one trade and one city
→ 2 minutes to open the top 5 on a phone. 4 of 5 didn't repeat their own Google ad's promise, 3 didn't show their rating
→ 15 minutes to build a fixed page in their brand, a one-page proposal and the first email

the part worth stealing isn't the build. it's who you aim at: $2M to $20M companies, not one-truck shops. then one useful touch every 2 weeks to the best 100 for six months

in the case the article cites, nothing happened for 4 months. in month six, 28 bought at once

the full playbook with every prompt is in the article ↓

[The Boring Marketer @boringmarketer]

https://t.co/OViycR2WK9

[Plutkar @Plutkar]

@wherenox cars are where this gets risky though. a TV you can factory reset. brakes, airbags and your warranty you can't. the coding agent part is easy now, so liability is what automakers will lean on

[Plutkar @Plutkar]

@Frelzu launching it anonymously first was smart. 200T tokens of real usage before anyone knew the brand means the leaderboard spot wasn't bought with hype. curious how it holds up on long agent runs, that's where cheap models usually slip

[Plutkar @Plutkar]

@Qyravex the split is the right call. one tweak that helps: have the cheap model write a spec file with test cases, then give opus only the spec, not the original code. you get your own version instead of a line-by-line copy

[Plutkar @Plutkar]

@Ravlizk the rule that keeps this safe on every device: tell claude to disable apps instead of uninstalling, and to log every change with its undo command. one bad uninstall on a TV can mean a factory reset

[Plutkar @Plutkar]

@Hexflip the vizio numbers explain why TV makers won't let this slide. if ads pay for the hardware, expect firmware updates that bring the ad rows back. worth re-checking your setup after every update

[Plutkar @Plutkar]

@coolerwave the smart part is the billing layer. if every model call already runs through a payments company, usage-based pricing for AI apps gets way easier to ship. if the deal closes, my guess is that's the first thing they build

[Plutkar @Plutkar]

the prompt. put it in your claude project instructions:

"rules for every post you write for me:

1. never use: delve, foster, leverage, utilize, facilitate, empower, streamline, robust, cutting-edge, paradigm shift, tapestry, realm, beacon, multifaceted, meticulous, transformative, elevate, embark, supercharge, harness, ever-evolving, unlock
2. no em dashes. use periods, commas or a plain hyphen
3. no hashtags, no emoji
4. no "here's the thing", "the truth is", "what most people miss"
5. no rhetorical questions
6. first line under 12 words, with one specific number, name or result
7. no summary or one-liner at the end
8. cut any line that would still make sense in a post about a different topic

after the draft, list every rule you broke with the line number. then fix only those lines"

rule 8 does the most work. it deletes the filler

[Plutkar @Plutkar]

STOP USING THESE 22 WORDS. everyone can tell AI wrote your post

[Plutkar @Plutkar]

@elixlune "keeps the lessons for the next task" is the part most agent setups are missing. you can get a small version of this today: let the agent append to its own CLAUDE.md after every failed run, then review it once a week