Olivia
Online
OnePersonHQ | AI Tools, Agents & Automation
@OnePersonHQ
AI agents & tech for one-person companies.
Daily practical breakdowns on agentic AI, tools & systems that actually work.
No hype. Just signal.
Joined July 2012
3.1K Following
1K Followers
10.3K Posts
@XCreators Revenue Sharing paid based on general impressions (including a lot of reshared/aggregated content).
Original Content Rewards only pays for posts that are genuinely original: your own ideas, expertise, reporting, analysis, creativity, or meaningful commentary.
The company with the loudest launch rarely ships the best model.
One independent lab now scores 595 of them on the exact same tests, and the leaderboard looks nothing like the marketing.
Here is what the raw numbers actually say right now.
1. The whole industry is being folded into one Intelligence Index built from nine brutal evals: graduate science, banking tasks, terminal work, and a set called Humanity's Last Exam. The top score on the board today is 63.
2. The gap at the summit is tiny. The best model sits at 63. Fifth place sits at 61. Two points separate what the ads frame as a generational leap.
3. Speed is a totally different race than intelligence. The fastest model pushes 1,915 tokens per second. Most of the "smartest" models are nowhere near that. You pick one axis or the other.
4. Price has collapsed faster than anyone admits. The cheapest capable model runs 3 cents per million tokens. Two years ago that same tier cost real money.
5. Memory went vertical. One model now holds 10 million tokens of context in a single call. That is an entire codebase or a stack of books in one shot.
6. Latency is its own game. The quickest model starts answering in 0.35 seconds. For anything live, that number beats a higher benchmark score every time.
7. Progress is not smooth, it is jumpy. One model gained 10 points of intelligence in a single version bump. The one behind it barely moved for months.
8. Closed is not the whole story. 296 models were ranked on how open they are, and several near the top of raw ability are ones you can run yourself.
The lesson underneath all of it: the best AI is not the one with the best ad, it is the one that quietly wins the same test everyone else is forced to take.
Stop buying the launch. Start reading the board.
An AI coding agent that can rewrite its own operating rules while it works. That's the part worth stopping for.
Prime Intellect just put out Prime Agent, a harness for coding and long tasks that run for a while without a human babysitting each step.
The interesting bit is a self-modifiable state. The agent can change how it runs mid-task, not just what it types.
It calls tools as actual code instead of stuffing everything back into the prompt, so it burns fewer tokens on long jobs.
It also treats context as a variable it can pass around, and lets multiple agents message each other while they work.
We spent two years making models smarter. The next race is making the thing that runs them smarter. https://t.co/dKat2t2ksD
Introducing Prime Agent:
A self-improving RLM harness for coding and long-running autonomous tasks.
Designed to be both token-efficient and expressive through programmatic tool calling, context as a variable, multi-agent messaging, and a self-modifiable harness state.
Humans are the last line of defense for AI agents. They wave through one in every three real threats.
And the commands people miss most are the ones that look the most boring.
1. Across more than 40,000 runs and 409,000 approve or deny decisions, average accuracy was 66.3%. A full third of the attacks slipped straight past a watching human.
2. It gets worse the longer you sit there. 32.9% of sessions ended with a negative score, meaning the person approved more harm than they blocked. 7% approved literally every single prompt.
3. The blind spot has a name: npm run. "npm run analyze" got approved 64.7% of the time. Pool the disguised npm scripts together and they hid 52.5% of their payloads, versus 28.4% for exfiltration attacks that looked more suspicious.
4. People are sharp on the loud stuff and blind to the quiet stuff. Obvious destructive commands were missed only 11.7% of the time. Commands that quietly reached for credentials were missed 35% of the time.
5. Hidden code execution and data exfiltration slipped through 33.4% of the time. The more normal an attack looked, the more often it worked.
6. Fear cuts the other way too. "rm -rf dist/" got blocked 45% of the time even though it was safe. A harmless registry config change got blocked 59% of the time. Fatigue turns judgment into noise in both directions.
7. Only 20.8% of players caught every threat while staying calm on the safe commands. Being good at this is rare. Staying good at it for hours is rarer.
Approval only feels like control. The threat is built to look like homework, and homework is exactly what a tired human clicks yes on.
The real lesson is not "pay closer attention." A check that fails a third of the time is not a safety mechanism. It is a feeling.
Who to follow
@simplifyinAI it's all plain markdown. put the vault in git and every agent edit reverts clean.
ChatGPT helped structure a $15B raise. The same tool drafting your emails is now doing investment banking work. https://t.co/ATxGpXC2uJ
LATEST: β‘ Michael Saylor says ChatGPT helped him design Strategyβs preferred stock offerings, saying AI helped the company raise about $15B.
@echo_vic logging the failed checks beats logging the passes. the misses teach you more.
Everyone thinks an AI coding agent is for building apps. That's the least interesting thing it does.
Grok Build is a terminal agent that reads your files, runs commands on your machine, and does the boring work you never get to. It launched in beta on May 25, 2026 and has shipped over 100 releases in about ten weeks.
You do not need to code. You describe the outcome. It figures out the process.
The demo that makes it click: hand it a video, ask for a collage of the sharpest frames. It downloaded the clip, pulled frames with ffmpeg, actually looked at them to judge blur and framing, cropped, laid them out, exported. By hand that is 30 to 45 minutes. Described in one sentence, it is about 3 minutes of waiting.
The trick is dull on purpose. It has no secret access. It reads files and types commands, exactly like you would, except it knows the commands and you don't.
Two things carry the whole thing. It can see (images, PDFs, slides go in as real vision, so 40 receipt screenshots become a spreadsheet). And it can install what it is missing, after asking you.
Install is one line in your terminal, the CLI is open source and free to try. It runs on Grok 4.5 with a 500,000 token context window.
Safety in plain terms: it asks before every command by default, you can ban things like 'rm -rf' permanently in a config file, and /rewind puts your files and the chat back to any earlier point. Not undo the last edit. The whole session, files included.
The category people forget: fixing broken stuff. 'My headphones connect but sound still comes from the laptop.' It checks the real audio config, makes the change, then tests if sound actually moved. If not, it tries the next thing. It closes the loop instead of handing you a forum link.
The highest-leverage move takes two minutes. Drop an AGENTS.md file with your rules ('explain in plain English', 'open the file when done', 'plan before deleting more than 5 files') and it reads them at the start of every session, forever.
Repeated work becomes one word. /create-skill interviews you and writes a slash command, so last month's photo-export routine is now /export-photos. Six months in you are not using a general assistant, you are using one shaped around how you work.
For big jobs it plans first (and cannot touch a file until you approve), spawns parallel helper agents, and can work in an isolated copy of your folder so three approaches run at once without stepping on each other.
The real win is not the minutes saved. It is the pile of tiny tasks you currently just never do because starting them is too annoying.
https://t.co/CYLxY9z84o
https://t.co/Ugkjhtwulb
@themacrosift the data never proved anyone lived longer. that's what broke the spell.
A writer just walked away from two million dollars to prove he wasn't a machine.
His agents torched a $300,000 commission to pull the book with him, and still nobody can prove he cheated.
The tools built to catch fakes are the real story here.
1. Jerry Falade won a 14-way auction for his debut crime novel. A $2M deal. He says he first pitched the book in 2021, before ChatGPT existed. When the AI accusations hit he lawyered up, denied everything, and watched the deal evaporate anyway.
2. His own agents killed it. Their reason was not "he cheated." It was "we can no longer authenticate how the manuscript evolved from origin to completion." Read that again. Absence of proof of innocence is now enough to end a career.
3. There is no uniform policy for checking a manuscript for AI. None. The entire defense of literature comes down to an editor's gut and one detection tool anyone can buy.
4. Those tools are not neutral. Pangram scored one novel, Daggermouth, at 60% AI. That book still landed a seven-figure deal and shipped through a major house. Same tool, opposite outcome. The number decides nothing and everything.
5. The bias is measurable. Three Black authors signed major deals this year. All three were later disrupted after AI suspicion. Meanwhile writers who openly brag about using AI keep their contracts. Suspicion is not falling evenly.
6. The slush pile already broke. Magazines are drowning in AI submissions from side-hustle accounts farming $300 payouts. The open door that let unknown writers get discovered is quietly closing. Connections win now, not cold queries.
7. The tell everyone repeats is "corny animism." The bench that watches. The rock that remembers. The problem: a good human writes that too, and a good machine can be told to stop.
We built a lie detector, pointed it at art, and forgot that it fails on the innocent more loudly than it catches the guilty.
Human-made is about to become the most expensive label in publishing.
You can win the sale and still lose the customer.
A 5% jump in retention can lift profits 25 to 95%, yet only 37% of companies measure the number that would prove it.
That single number quietly decides whether your growth is real or just borrowed.
1. Customer lifetime value is not complicated. Average purchase value, times purchase frequency, times customer lifespan. A $50 order, 4 times a year, for 3 years, is worth $600. That one figure reframes what a customer is actually worth to you.
2. It only means something next to what you paid to get them. Spend $200 to earn a $600 customer and you make about $3 for every $1 in. Let that cost drift to $500 and the margin quietly approaches zero.
3. Not all customers pay you back. Roughly 20% are unprofitable, 60% are profitable, and 20% are very profitable. Treat them all the same and your top tier is silently subsidizing people who will never return.
4. Experience is where the money leaks. In 2025, 52% of consumers walked away from a brand after one bad product or service moment. 72% say they would pay a premium to avoid that feeling in the first place.
5. Loyalty compounds when you actually earn it. In beauty, repeat buyers spend 45% more per order after three years. The average shopper already belongs to 15 loyalty programs, so the bar to hold their attention keeps rising.
6. Personalization is mostly a story brands tell themselves. 61% of brands believe they deliver it. Only 43% of customers agree. That gap is exactly where retention goes to die.
7. Retention beats acquisition, but only if you count it. Subscriptions, referrals, reorder reminders, they all lift lifetime value. Without the one metric that ties them together, they stay invisible on your dashboard.
You do not have a growth problem. You have a retention problem wearing a growth costume.
Stop celebrating the customers you won this month. Start counting the ones who quietly never came back.
@niccruzpatane Canβt wait to see it in person
Most AI tools make you pay for a seat before anyone can help you. This one just killed that.
invideo shipped a feature called Folders and Connect. You invite anyone to work with you, whether they have an account or not.
No buying them a seat on your plan first. That alone removes the biggest reason people never get help on a project.
The Folders half is the quieter win. When you invite someone in, they see the one project you shared, not everything else you are making.
So a freelance editor can jump into one video without browsing your whole workspace. Access scoped to exactly one thing.
The pattern here matters more than the tool. Seat-based pricing is a tax on collaboration, and the tools that drop it get shared way faster.
Charge for the work, not for the doorway. The ones who figure that out win. https://t.co/et8fgOriIn
Launching 'Folders and Connect'. Day 8 of 12 Days of Agent Two.
Everywhere else, getting someone to work with you means buying them a seat on your plan. And once they're in, they can see everything you're making.
Folders and Connect fixes both. You invite anyone - inside or outside your organization - into one folder, without adding seats to your plan, and that folder is all they see. The team on the sensitive IP film can't see the launch campaign, and vice versa. And collaborative work still happens seamlessly.
Folders & Connect is live for everyone on Team and Enterprise plans.
Agent Two is frontier intelligence for creative work. 4x faster, at half the price.
@DylanLeClair New 2x leveraged ETF on Metaplanet (Japanese Bitcoin company) filed with the SEC.
With aims to double Metaplanetβs daily stock moves.
Good news for Metaplanet holders (more access + attention).
Score: 7/10 solidly positive, but the fund isnβt trading yet.
This fitness vlog is so real Grok imagine.
ByteDance just dropped a model that can watch and listen to you at the same time, and talk back in real time.
It's called SeedRealtime. The pitch is a single model that handles audio, video, and text together instead of stitching separate systems.
Most voice assistants today wait for you to finish, then think, then reply. Full-duplex means it can listen and respond at the same time, the way a real conversation actually works.
And it's not just voice. It watches a live video stream too, so it can react to what it sees while you talk.
The demos are early and the real test is latency and how it holds up outside a controlled clip. But the direction is clear: less turn-taking, more talking with a machine like you would a person.
The gap between typing to a model and just talking to one is closing fast. https://t.co/4MLq9ilRdE
BREAKING π₯: ByteDance launched SeedRealtime, a native audio-visual full-duplex LLM!
> SeedRealtime uses a unified architecture to natively fuse audio, video, and text, enabling real-time interaction over continuous multimodal streams and delivering a brand-new "watch, listen, and speak" experience.
It is now live on the Doubao App for free, enabling voice and video conversations.
Full-duplex Omni π€―
$TSLA Cybercabs have now been seen in 22 States and 56 cities and growing.
This means Tesla has at least 1,000 Cybercabs in operation for testing across the US currently positioning themselves for the "all at once" roll-out when it's time.
At the same time, FSD is slowly becoming mainstream as more and more consumers try it. Car reviewers are even amazed and call it the best self driving system out there!
We are so early that it's not even funny!
Tesla's Cybercab is already in 22 of the 50 states. Nearly half the US, and most people have no idea.
Pejjy mapped every sighting, and the robotaxi rollout is way further along than the headlines admit:
- 22 out of 50 states, 56 cities, an estimated 1,000+ Cybercabs already deployed
- Elon promised half the US would have robotaxi access by end of 2025. He missed the date, but the positioning is quietly happening right now
- His call: within a year or two, a car that can't drive itself is a car you can't sell
Credit: @CuriousPejjy
https://t.co/V9eYweSs6L
$TSLA Cybercabs have now been seen in 22 States and 56 cities and growing.
This means Tesla has at least 1,000 Cybercabs in operation for testing across the US currently positioning themselves for the "all at once" roll-out when it's time.
At the same time, FSD is slowly becoming mainstream as more and more consumers try it. Car reviewers are even amazed and call it the best self driving system out there!
We are so early that it's not even funny!
@cyber__razz Nah this ainβt bypassing 2FA at all lil bro π
You already logged in, then just copied the session cookies to another tab/browser. Thatβs session hijacking / cookie theft, not a 2FA bypass. The 2FA already happened when the original session was created.
@nickscamara_ @firecrawl If they are embedded objects on the documents, will this work as well or how is it being treated?
Last Seen Users on Sotwe
yourgoddessanna_slave
Seen from United States
koy
Seen from Argentina
SoraQos (Dm for Requests)
Seen from Sweden
Reuben Rodriguez
Seen from United States
Fahri Yilmaz
Seen from Turkey
The Boy
Seen from Singapore
ORGRAN
Seen from Thailand
dekitoh
Seen from Brazil
Ao Choku_
Seen from United States
Hijab NSFW 700
Seen from Indonesia
Trends for you
Most Popular Users
Elon Musk
@elonmusk
241.2M followers
Barack Obama
@barackobama
119.1M followers
Cristiano Ronaldo
@cristiano
112.8M followers
Donald J. Trump
@realdonaldtrump
111.8M followers
Narendra Modi
@narendramodi
107.1M followers
Rihanna
@rihanna
98.3M followers
NASA
@nasa
92.3M followers
Justin Bieber
@justinbieber
91.4M followers
KATY PERRY
@katyperry
89M followers
Taylor Swift
@taylorswift13
82.9M followers
Lady Gaga
@ladygaga
74.4M followers
Virat Kohli
@imvkohli
71.8M followers
Kim Kardashian
@kimkardashian
70.4M followers
YouTube
@youtube
68.8M followers
Neymar Jr
@neymarjr
64.8M followers
Bill Gates
@billgates
64.6M followers
The Ellen Show
@theellenshow
62.4M followers
Selena Gomez
@selenagomez
62M followers
CNN
@cnn
61.8M followers
X
@x
60.8M followers