Pinned Tweet
Anthropic CEO:
"This is a super weapon. You should have to own a gun license to use it. Please don't release this."
In his interview, Dario Amodei explains why he is keeping the newest model, Mythos, hidden from the public.
Autonomously goes through all the links of the cyber kill chain. 271 new vulnerabilities found in Firefox. Enormous commercial losses from refusing a public release.
He admits that today's cyber safeguards can still be jailbroken too easily, and he is waiting for a time when the internet ecosystem becomes more secure.
The most responsible voice in AI, or a brilliant PR strategist who turned a product delay into the ultimate marketing?
$100 a month into NIKE for 39 years
> invested: $47,200
> now: $761,748
> peak: $3.07m (nov 2021)
boring consistency beats timing the market
but nobody told nke the last 5 years
dca or lump sum?
If you’re on an OpenAI Pro plan, check your limits right now 👀
they quietly doubled across the board:
5x → 10x
10x → 20x
25x → 50x
Pro 200 keeps its limits, nothing cut
Dots has its own separate usage pool
surplus compute = a model launch needs room to breathe
GROK ULTRA is being given away for FREE
right now, you can get a 7-day subscription for $0
the best opportunity to test Grok and Grok Bot
don't miss this opportunity while X is being so generous
gkrovso retweeted
this is the new meta for promoting apps on tiktok
-> 188.3k followers on tiktok
-> 9.4M likes
it's arrow, a workout and diet planning app, and the bio says "get your diet & training plan on the arrow app with 1M+ others"
the format:
-> clean slideshows, white background, one bright green accent
-> a faceless anatomy figure with the muscles colored in
-> titles like "ranking every bicep exercise, from worst to best" and "supplement tier list"
why it works:
- people save these to open at the gym, so every save is someone planning their next workout
- and planning workouts is exactly what the app does
- rankings and tier lists make people comment to argue
for a cooking app, "ranking every protein by cost per gram"
for a language app, "every tense ranked from easiest to hardest" with one color per tense
want this running for your app? DM me "UGC"
my GROK Agents worked all night
the account had $50 when i went to bed
when i woke up it showed $2,145
finally enough money to survive the rest of the month
i opened the logs because there was absolutely no chance i was trusting that number
> SCOUT scanned 38 prediction markets and pulled the resolution rules price history expiry and every source connected to the event
> ROUTER grouped them by catalyst and removed anything that would resolve before the research was finished
> CRITIC killed 31 setups because the spread was ugly the orderbook was empty or the edge disappeared after fees
> VERIFIER caught one market where the headline implied one thing but the actual resolution text said another
>OPERATOR kept only two positions after every risk condition passed
the weird part is that none of them tried to force a trade
they spent most of the night proving why almost every opportunity was bad
basically the exact opposite of how i trade
now i can finally afford groceries rent and maybe one coffee
unless the agents decide they deserve a performance fee
i connected 72 agents to one task queue because for the first hour it looked less like intelligence and more like a crowded train station
the first overnight forecast was $378.84
by morning only 14 agents still had jobs and the bill had fallen to $26.85
the router caught 39 workers running toward tasks another worker had already finished
another 19 kept writing explanations that no downstream step requested
the remaining 14 received one role one input and one visible exit
> they processed 24,739 tasks
> median latency dropped from 36 seconds to 6.6 seconds
> accepted output climbed from 74 percent to 95 percent
> retries fell from 5,264 to 679
> the queue went from 7,313 items to ZERO at 6:13 am
the useful change was not a larger model
it was a dispatcher deleting duplicate work before the agents could race toward it
i spent the evening building an army and the morning removing 58 access cards
the last workers are moving faster now that nobody is standing in the doorway
i connected 72 agents to one task queue because for the first hour it looked less like intelligence and more like a crowded train station
the first overnight forecast was $378.84
by morning only 14 agents still had jobs and the bill had fallen to $26.85
the router caught 39 workers running toward tasks another worker had already finished
another 19 kept writing explanations that no downstream step requested
the remaining 14 received one role one input and one visible exit
> they processed 24,739 tasks
> median latency dropped from 36 seconds to 6.6 seconds
> accepted output climbed from 74 percent to 95 percent
> retries fell from 5,264 to 679
> the queue went from 7,313 items to ZERO at 6:13 am
the useful change was not a larger model
it was a dispatcher deleting duplicate work before the agents could race toward it
i spent the evening building an army and the morning removing 58 access cards
the last workers are moving faster now that nobody is standing in the doorway
Tibo just confirmed we’re getting another Codex + ChatGPT usage reset today at 6pm PST
these resets are basically happening every other day now
at this point there’s no time to sleep, you either use your limits or watch them reset again
Grok Bot is the FIRST AGENT product that got the boring part RIGHT
every other agentic tool treats persistence as an afterthought, session dies, context resets, you start over
Grok Bot's answer is a shared persistent vm that survives across tasks and across bots on the same account
that's not a small design choice
it means a research bot can hand off to a writer bot mid-task without you re-explaining context, because they're literally running on the same machine with the same authenticated state
the tradeoff is real (shared vm means shared blast radius, worth designing role scopes carefully)
but that's a config problem, not an architecture problem. and config problems get solved
most "ai agents" today are still glorified chatbots with a loop bolted on
this is the FIRST ONE that actually feels like infrastructure
CEO's from Coinbase, Ripple, Gemini, Robinhood, Polymarket and Kalshi sit on the new CFTC innovation committee
traditional finance chiefs from CME, NASDAQ, ICE and DTCC join them
Bessent and Lutnick may also show per sources
thursday CFTC session tackles federal market structure and the digital asset clarity act
Trump attendance may hinge on accepting tighter ethics rules around his own crypto ties
real policy signal or just a photo op before the senate vote?
gkrovso retweeted
Grok 4.6 is out
in terms of intelligence, coding and agentic capabilities, it's already roughly on the level of Sol/Fable
but the most interesting part here isn't even that, it's the price
- GPT-5.6 Sol: $5 in / $30 out
- Claude Fable 5: $10 in / $50 out
- Grok 4.6: $2 in / $6 out
that means Grok is almost 5x cheaper than Sol on output and more than 8x cheaper than Fable
the model is also more token-efficient than GPT
we'll see the full charts once they come out
> Grok 4.6 scored 61 on the AA Intelligence Index, the same as GPT-5.6 Sol Max and just 1 point below Claude Fable 5 Max
> on DeepSWE 1.1, the score went from 54% with Grok 4.5 High to 65.9%
the model is built for long-running agentic tasks, agentic coding and creating interactive and visual projects
it can take complex tasks all the way to completion, build a first version of an app from a general idea, improve it based on feedback and test it on its own
there's also a new xhigh reasoning mode
already available in Cursor, Grok Build, API, OpenRouter, Vercel and Cloudflare
and in Cursor and Grok Build, they even doubled the limits for the first week
intelligence is getting cheaper unexpectedly
Grok 4.6 is out
in terms of intelligence, coding and agentic capabilities, it's already roughly on the level of Sol/Fable
but the most interesting part here isn't even that, it's the price
- GPT-5.6 Sol: $5 in / $30 out
- Claude Fable 5: $10 in / $50 out
- Grok 4.6: $2 in / $6 out
that means Grok is almost 5x cheaper than Sol on output and more than 8x cheaper than Fable
the model is also more token-efficient than GPT
we'll see the full charts once they come out
> Grok 4.6 scored 61 on the AA Intelligence Index, the same as GPT-5.6 Sol Max and just 1 point below Claude Fable 5 Max
> on DeepSWE 1.1, the score went from 54% with Grok 4.5 High to 65.9%
the model is built for long-running agentic tasks, agentic coding and creating interactive and visual projects
it can take complex tasks all the way to completion, build a first version of an app from a general idea, improve it based on feedback and test it on its own
there's also a new xhigh reasoning mode
already available in Cursor, Grok Build, API, OpenRouter, Vercel and Cloudflare
and in Cursor and Grok Build, they even doubled the limits for the first week
intelligence is getting cheaper unexpectedly
Grok 4.6 is out
in terms of intelligence, coding and agentic capabilities, it's already roughly on the level of Sol/Fable
but the most interesting part here isn't even that, it's the price
- GPT-5.6 Sol: $5 in / $30 out
- Claude Fable 5: $10 in / $50 out
- Grok 4.6: $2 in / $6 out
that means Grok is almost 5x cheaper than Sol on output and more than 8x cheaper than Fable
the model is also more token-efficient than GPT
we'll see the full charts once they come out
> Grok 4.6 scored 61 on the AA Intelligence Index, the same as GPT-5.6 Sol Max and just 1 point below Claude Fable 5 Max
> on DeepSWE 1.1, the score went from 54% with Grok 4.5 High to 65.9%
the model is built for long-running agentic tasks, agentic coding and creating interactive and visual projects
it can take complex tasks all the way to completion, build a first version of an app from a general idea, improve it based on feedback and test it on its own
there's also a new xhigh reasoning mode
already available in Cursor, Grok Build, API, OpenRouter, Vercel and Cloudflare
and in Cursor and Grok Build, they even doubled the limits for the first week
intelligence is getting cheaper unexpectedly