anthropic is secretly testing something called Conway and it's actually insane
so basically you give the AI a task and go to sleep, wake up - everything is done, not joking
this is NOT chatgpt where you sit there prompting like a monkey, Conway runs 24/7 IN THE BACKGROUND, it opens browsers by itself, clicks buttons, fills out forms, makes decisions on its own, you literally say "monitor competitor prices" and it JUST DOES IT, every single day, without you
if it's not sure - it asks, if it's sure - it executes and reports back
anthropic basically built the first AI that doesn't answer questions but actually WORKS, like an intern who never sleeps, never eats, never takes PTO and never messages you on slack "hey is this actually my task?"
the leak came from security researchers lmao they exposed the whole architecture and it's a full blown OS for AI agents with triggers, webhooks, persistent memory between sessions
so basically you give the AI a task and go to sleep, wake up - everything is done, not joking
this is NOT chatgpt where you sit there prompting like a monkey, Conway runs 24/7 IN THE BACKGROUND, it opens browsers by itself, clicks buttons, fills out forms, makes decisions on its own, you literally say "monitor competitor prices" and it JUST DOES IT, every single day, without you
if it's not sure - it asks, if it's sure - it executes and reports back
anthropic basically built the first AI that doesn't answer questions but actually WORKS, like an intern who never sleeps, never eats, never takes PTO and never messages you on slack "hey is this actually my task?"
the leak came from security researchers lmao they exposed the whole architecture and it's a full blown OS for AI agents with triggers, webhooks, persistent memory between sessions
❤10🔥2
OpenAI is losing to Anthropic and panicking
their answer - a superapp: ChatGPT + Codex + Atlas browser in one window, full agentic loop, Brockman personally leading the build
the problem - Anthropic already has this working, plus Conway is coming, OpenAI's entire bet is on a new model called "Spud", if it flops the migration to Claude only accelerates
their answer - a superapp: ChatGPT + Codex + Atlas browser in one window, full agentic loop, Brockman personally leading the build
the problem - Anthropic already has this working, plus Conway is coming, OpenAI's entire bet is on a new model called "Spud", if it flops the migration to Claude only accelerates
last story from NYT
guy from LA, zero coding, $20k, a bunch of AI tools - built a telehealth company selling weight loss drugs, $401M revenue in year one, 2 employees total (him and his brother), NYT called him the AI success story of 2026
behind the curtain
"$1.8B company" - no investors, no valuation, just a revenue projection, real profit is $65M
"2 employees" - hundreds of people work behind the scenes through outsourced medical platforms
FDA sent a warning letter for fake AI-generated before/after photos, affiliate pages had fake doctors (one was literally a gospel musician's old page), 3 lawsuits for spam, Techdirt called it "NYT got played by a scam"
takeaway
AI removes the friction between your idea and the market - if the idea is good you scale fast, if the idea is fraud you also scale fast
the tools are the same for everyone, the difference is what you build with them
story from NYT: https://www.nytimes.com/2026/04/02/technology/ai-billion-dollar-company-medvi.html
guy from LA, zero coding, $20k, a bunch of AI tools - built a telehealth company selling weight loss drugs, $401M revenue in year one, 2 employees total (him and his brother), NYT called him the AI success story of 2026
behind the curtain
"$1.8B company" - no investors, no valuation, just a revenue projection, real profit is $65M
"2 employees" - hundreds of people work behind the scenes through outsourced medical platforms
FDA sent a warning letter for fake AI-generated before/after photos, affiliate pages had fake doctors (one was literally a gospel musician's old page), 3 lawsuits for spam, Techdirt called it "NYT got played by a scam"
takeaway
AI removes the friction between your idea and the market - if the idea is good you scale fast, if the idea is fraud you also scale fast
the tools are the same for everyone, the difference is what you build with them
story from NYT: https://www.nytimes.com/2026/04/02/technology/ai-billion-dollar-company-medvi.html
🔥4
anthropic released a model so powerful the US government held an emergency meeting with Wall Street banks
claude mythos can autonomously find and exploit vulnerabilities in any browser and OS, it chains exploits, escalates privileges, moves laterally, like a hacking team that never sleeps
this is NOT some chatbot finding bugs in your code, mythos BREAKS INTO SYSTEMS on its own, you point it at a target and it JUST GOES
treasury secretary and fed chair personally called in bank CEOs to warn them, first time ever a government treats an AI release as a threat to the financial system
market reacted instantly, cloudflare down 12%, palantir down 7%, crowdstrike down 6%, michael burry said anthropic is "eating palantir alive"
only ~50 organizations got access, everyone else locked out, anthropic themselves said capabilities are "too high" for public release
we went from "AI writes your emails" to "AI triggers emergency government meetings" in about a year
claude mythos can autonomously find and exploit vulnerabilities in any browser and OS, it chains exploits, escalates privileges, moves laterally, like a hacking team that never sleeps
this is NOT some chatbot finding bugs in your code, mythos BREAKS INTO SYSTEMS on its own, you point it at a target and it JUST GOES
treasury secretary and fed chair personally called in bank CEOs to warn them, first time ever a government treats an AI release as a threat to the financial system
market reacted instantly, cloudflare down 12%, palantir down 7%, crowdstrike down 6%, michael burry said anthropic is "eating palantir alive"
only ~50 organizations got access, everyone else locked out, anthropic themselves said capabilities are "too high" for public release
we went from "AI writes your emails" to "AI triggers emergency government meetings" in about a year
cursor just killed the code editor and rebuilt it as an AI command center
you open cursor 3.0 - no file explorer, no tabs, just an agent orchestration panel where you launch unlimited AI agents in parallel across different repos
this is NOT "we added copilot to your IDE", the IDE is GONE, agents are the default view now
killer feature - Design Mode: Cmd+Shift+D, click any element in your app's UI, tell the AI what to change, done, no more describing buttons in chat like an idiot
agents review each other's code, run locally and in the cloud simultaneously, screenshot pages and click by visual coordinates when DOM fails
one dev running 10 agents approving PRs - that's the workflow now
we went from "AI autocompletes your line" to "AI runs your entire engineering team" in 18 months
you open cursor 3.0 - no file explorer, no tabs, just an agent orchestration panel where you launch unlimited AI agents in parallel across different repos
this is NOT "we added copilot to your IDE", the IDE is GONE, agents are the default view now
killer feature - Design Mode: Cmd+Shift+D, click any element in your app's UI, tell the AI what to change, done, no more describing buttons in chat like an idiot
agents review each other's code, run locally and in the cloud simultaneously, screenshot pages and click by visual coordinates when DOM fails
one dev running 10 agents approving PRs - that's the workflow now
we went from "AI autocompletes your line" to "AI runs your entire engineering team" in 18 months
stanford just dropped a 400 page report and the numbers are insane
ai models now match or BEAT human experts on economically valuable tasks, gpt-5.4 scored 83% on gdpval which is above expert level
92% of code in the US is now written with ai tools, 46% of new code is fully ai-generated, the industry went from "ai helps you code" to "ai codes for you" in one year
but here's the dark part - pwc says the gains are flowing to a SHRINKING circle of winners, meaning most companies and most workers are getting left behind while a small group prints money
us-china gap is widening, the us is pulling ahead on frontier models and china can't close the compute difference, morgan stanley warned that a breakthrough in the first half of 2026 will "shock" most people
we're not in the hype phase anymore, we're in the "if you're not using this you're already behind" phase
ai models now match or BEAT human experts on economically valuable tasks, gpt-5.4 scored 83% on gdpval which is above expert level
92% of code in the US is now written with ai tools, 46% of new code is fully ai-generated, the industry went from "ai helps you code" to "ai codes for you" in one year
but here's the dark part - pwc says the gains are flowing to a SHRINKING circle of winners, meaning most companies and most workers are getting left behind while a small group prints money
us-china gap is widening, the us is pulling ahead on frontier models and china can't close the compute difference, morgan stanley warned that a breakthrough in the first half of 2026 will "shock" most people
we're not in the hype phase anymore, we're in the "if you're not using this you're already behind" phase
❤8
anthropic just dropped opus 4.7 and basically admitted it's not even their best model anymore
released april 15. the upgrade isn't about benchmarks - it's that the model finally stops losing the plot mid-task and validates its own output before shipping
SWE Bench Pro - 64.3, follows instructions literally (your old prompts may break)
vision 3x better: images up to 2576px, XBOW visual pentest 98.5% vs 54.5% on 4.6 - computer-use agents finally see screenshots properly
new /ultrareview mode - strict code review, re-validates the answer before returning it
spicy part - anthropic openly said 4.7 is "broadly less capable" than their internal Mythos model that's been running with select partners for a month
so the public flagship is a crash test dummy for the safety filters they want to ship on Mythos - first time anthropic shipped "the best model we're allowed to give you" instead of "the best we have"
downside - eats more tokens, daily limits will evaporate faster
released april 15. the upgrade isn't about benchmarks - it's that the model finally stops losing the plot mid-task and validates its own output before shipping
SWE Bench Pro - 64.3, follows instructions literally (your old prompts may break)
vision 3x better: images up to 2576px, XBOW visual pentest 98.5% vs 54.5% on 4.6 - computer-use agents finally see screenshots properly
new /ultrareview mode - strict code review, re-validates the answer before returning it
spicy part - anthropic openly said 4.7 is "broadly less capable" than their internal Mythos model that's been running with select partners for a month
so the public flagship is a crash test dummy for the safety filters they want to ship on Mythos - first time anthropic shipped "the best model we're allowed to give you" instead of "the best we have"
downside - eats more tokens, daily limits will evaporate faster
❤5
openai codex desktop just ate the dev's entire machine - 3M weekly users, full computer-use now default on macOS
openai shipped a massive codex update this week - background computer-use on macOS means the agent takes over your whole desktop, not just your editor, clicks around apps, runs shell commands, edits files across repos while you sleep
3 million devs use it weekly already, and with the new local-cloud parallel execution you fire off 10 agents, go get coffee, come back to PRs waiting for review
the IDE is no longer where code happens - code happens everywhere the cursor can reach, and you're the PM of a swarm
openai shipped a massive codex update this week - background computer-use on macOS means the agent takes over your whole desktop, not just your editor, clicks around apps, runs shell commands, edits files across repos while you sleep
3 million devs use it weekly already, and with the new local-cloud parallel execution you fire off 10 agents, go get coffee, come back to PRs waiting for review
the IDE is no longer where code happens - code happens everywhere the cursor can reach, and you're the PM of a swarm
👨💻6
while everyone was losing their minds over o3 openai quietly dropped gpt-4.1 and almost nobody noticed
three versions: gpt-4.1, gpt-4.1 mini, gpt-4.1 nano, api only, NOT in chatgpt
this is not a reasoning model like o3, this is the workhorse - faster, cheaper, better at following instructions, built for code and multimodal tasks
openai is literally fighting a war on two fronts at the same time, reasoning models for hard problems and classic models for everything else, and they updated BOTH on the same day
developers are already saying 4.1 mini is the best price-to-quality ratio on the entire market right now
the quiet release that will change more than the loud one
three versions: gpt-4.1, gpt-4.1 mini, gpt-4.1 nano, api only, NOT in chatgpt
this is not a reasoning model like o3, this is the workhorse - faster, cheaper, better at following instructions, built for code and multimodal tasks
openai is literally fighting a war on two fronts at the same time, reasoning models for hard problems and classic models for everything else, and they updated BOTH on the same day
developers are already saying 4.1 mini is the best price-to-quality ratio on the entire market right now
the quiet release that will change more than the loud one
apple just nuked the entire vibecoding app store - pulled anything ($100M valuation), froze replit and vibecode updates, and the reason is exactly what you think it is
apple quietly yanked vibecoding platforms off the app store this week citing "longstanding rules" but everyone knows the real play - these apps let randoms spin up full ios products from a prompt, bypassing app review and the sacred 30% cut
anything was sitting at a $100M valuation with users shipping apps in an afternoon, replit had of devs on mobile, vibecode was eating the no-code market - all three hit the wall at the same time on the same week, not a coincidence
this is the first real platform-holder strike against the vibecoding economy - if you can't distribute what the agent builds, the whole "i vibed an app in 20 minutes" loop collapses at the last mile, and apple owns the last mile
the IDE war was always a distraction - the real war is distribution, and cupertino just reminded everyone who owns the pipes
apple quietly yanked vibecoding platforms off the app store this week citing "longstanding rules" but everyone knows the real play - these apps let randoms spin up full ios products from a prompt, bypassing app review and the sacred 30% cut
anything was sitting at a $100M valuation with users shipping apps in an afternoon, replit had of devs on mobile, vibecode was eating the no-code market - all three hit the wall at the same time on the same week, not a coincidence
this is the first real platform-holder strike against the vibecoding economy - if you can't distribute what the agent builds, the whole "i vibed an app in 20 minutes" loop collapses at the last mile, and apple owns the last mile
the IDE war was always a distraction - the real war is distribution, and cupertino just reminded everyone who owns the pipes
some frens of mine recently joined copytrade on polymarket, and they’re already adding great features for users
+ bots that I share with them from my clusters, as mentioned in my latest article https://x.com/LunarResearcher/status/2049141918521491606
if anyone wants to check out how copytrade works with these bots - feel free to give it a try
more to come, but remember nfa + dyor always
+ bots that I share with them from my clusters, as mentioned in my latest article https://x.com/LunarResearcher/status/2049141918521491606
if anyone wants to check out how copytrade works with these bots - feel free to give it a try
more to come, but remember nfa + dyor always
Forwarded from PolyFire
What this means for you:
1. All orders are now in pUSD
2. Expanded deposit options: USDT, USDC, pUSD now available
3. Speed has improved - we've optimized our bot
To celebrate, we've handpicked a few smart wallets so you can copytrade and stack $$$
0xce835747e39bf987de4c04052b1677e6633e5b97Win Rate: 95%
PnL: +$2,586
[Copytrade]
0xc6a28c25a8409f2d5bea1d5041355eec097fd076Win Rate: 56.67%
PnL: +$93,124
[Copytrade]
0xf2a5c9c77863715c2cc1a6d34670e8b75ba155d6Win Rate: 71.43%
PnL: +$5,283
[Copytrade]
Please open Telegram to view this post
VIEW IN TELEGRAM
cursor's bugbot now learns from YOU in real time, your reviewer mutates overnight
cursor flipped bugbot from batch retraining to continuous online learning this week, every thumbs-down ships into the model tomorrow
first mainstream ide agent that evolves while you sleep, you're not using a tool, you're labeling a dataset with every keystroke
sounds cute until your reviewer mutates faster than you can audit, today it nags about null checks, next week it silently greenlights sketchy sql because three randos clicked "like"
rlhf used to be a lab thing openai did with paid annotators, now cursor does it to you, unpaid, on your prod codebase
the ide stopped being a tool and became a feedback loop, you think you're reviewing the agent, the agent is reviewing you
cursor flipped bugbot from batch retraining to continuous online learning this week, every thumbs-down ships into the model tomorrow
first mainstream ide agent that evolves while you sleep, you're not using a tool, you're labeling a dataset with every keystroke
sounds cute until your reviewer mutates faster than you can audit, today it nags about null checks, next week it silently greenlights sketchy sql because three randos clicked "like"
rlhf used to be a lab thing openai did with paid annotators, now cursor does it to you, unpaid, on your prod codebase
the ide stopped being a tool and became a feedback loop, you think you're reviewing the agent, the agent is reviewing you
swap the LLM in one config line, Memory, skills, and tool wiring survive the run
the whole workspace packages up and boots on someone else's machine in one click
holaOS calls it Environment Engineering and after the article u can't unsee it
https://x.com/LunarResearcher/status/2050510337737154984
the whole workspace packages up and boots on someone else's machine in one click
holaOS calls it Environment Engineering and after the article u can't unsee it
https://x.com/LunarResearcher/status/2050510337737154984
👨💻2
new InfoFI for Polymarket/AI agents
standard campaigns with a reward pool
join now while there aren't many people yet
avg ~$50 per post for accounts in the leaderboard
registration link
standard campaigns with a reward pool
join now while there aren't many people yet
avg ~$50 per post for accounts in the leaderboard
registration link
new article about quant machine
https://x.com/LunarResearcher/status/2056001315331784841
maybe so interesting theme now
https://x.com/LunarResearcher/status/2056001315331784841
maybe so interesting theme now
X (formerly Twitter)
Lunar (@LunarResearcher) on X
Claude + Polymarket: How I Built a Quant Machine That Turns $200 Into $12,000
new market with f1 monaco grand prix
lately, I've gotten into racing as a hobby
overall, the bet on other drivers are pretty, in my opinion
i'll place a bet, just for fun watching
if anyone's interested https://x.com/LunarResearcher/status/2062159692726509685
lately, I've gotten into racing as a hobby
overall, the bet on other drivers are pretty, in my opinion
i'll place a bet, just for fun watching
if anyone's interested https://x.com/LunarResearcher/status/2062159692726509685
X (formerly Twitter)
Lunar (@LunarResearcher) on X
Formula 1. Monaco Grand Prix 2026 Winner Market Research
Choice Markets - a fully on-chain UGC prediction market - gives Kimi Antonelli 31%, Charles Leclerc 27%, and Any Other Driver 42%
Market link: https://t.co/W92DhdyAc9
Let's look at the market:
Kimi…
Choice Markets - a fully on-chain UGC prediction market - gives Kimi Antonelli 31%, Charles Leclerc 27%, and Any Other Driver 42%
Market link: https://t.co/W92DhdyAc9
Let's look at the market:
Kimi…