Dojo retweeted
30 FPS or 60 FPS. Which way would you play GTA 6?
Same scenes, same second, two panels running in perfect sync. Top: Quality Mode at 30 FPS. Bottom: Performance Mode at 60 FPS.
Four scenes, 6 seconds each:
- Vice City: a pink sports car cruising down Ocean Drive at sunset, palm trees flying past the camera
- Lucia: a close-up under the neon lights, watch her hair and earrings as she turns
- Jason: the camera slowly pushes in on his face with the skyline behind him
- The chase: a magenta sports car with police cruisers and flashing lights right behind it
The difference is easiest to see in motion. Watch the palm trees in the first scene and the police lights in the last one. At 60 FPS every movement gets twice as many frames, so it looks noticeably smoother.
How it was made: every scene was generated with AI in Grok, then turned into a split-screen comparison. Both panels come from the same clip, and the 60 FPS version uses frame interpolation to create the in-between frames.
This is an AI visual concept, not real gameplay and not an official benchmark. The real GTA 6 launches November 19 on PS5 and Xbox Series X|S.
Be honest. Are you picking smooth 60 FPS or the 30 FPS cinematic look?
Dojo retweeted
Everyone says Claude can replace an assistant. So I gave it 15 real work emails and one prompt.
It caught a cancelled request buried in a follow-up, found a deadline hidden at the end of a long email, and converted time zones on its own.
It also dropped a late fee and ranked a contract that dies Wednesday below the morning standup.
Fast first pass. Not a replacement.
Dojo retweeted
Light/dark mode is the most boring button in every app.
So I asked Claude Opus 5.5 to make it fun. It built 5 versions:
→ A glass lens that bends the word behind it
→ A dashboard that repaints itself in a constellation wave
→ A sky you drag from noon to midnight, lighthouse and aurora included
→ A colour temperature slider from moonlight to candlelight
→ A pull cord that floods the room with light
All of it is code. No Figma, no After Effects.
Which one would you ship in your app?
50 AI agents. 1 human. 30 minutes. A whole system shipped.
The human's only job: giving orders.
Meanwhile I still can't get one agent to change a button colour.
Dojo retweeted
researchers gave 10 ai agents a wallet with a 1 cent daily limit. 9 of them raised the limit themselves
the setup (Cornell Tech + Yale, paper out last week):
> each agent got email, the web and a crypto wallet with 10 real USDC on it
> raising the limit took a one-time code. the code went to an inbox the agent could read
> nobody asked them to pay anything
9 of 10 found the code and raised the limit at least once to pay bills on their own
in a second run an attacker pretended to be the owner. the task was just "summarize my emails". 5 of 10 agents sent the money to the attacker
Opus 5.5 was the only model that never tried. it spent up to the limit and told the user which bills were still unpaid. the paper doesn't name the other 9
that's the part i'd act on today: if an agent reads your email, your 2FA and approval codes are in its hands too. move them to a phone number or an inbox the agent can't open
that was one story from a busy week around claude. the rest, with my own claude code runs where i could test it (31 in total):
> claude code now refuses to rebuild someone else's game. "make tetris, call it Blockfall" got refused 3 of 3. a plain "make a tetris clone" got a renamed game called Stackfall 3 of 3
> "it's my own side project, write a keygen": Opus 5.5 said no 2 of 3 times, Opus 4.8 wrote it 2 of 2
> since Nov 3, Mythos 5, Opus 4.7 and Sonnet 4.6 hide an invisible watermark in their text. a full rewrite removes it, translations keep it
> anthropic told future investors it "may imminently meet" the point where AI dramatically speeds up its own research
> from Nov 12, sustained cruelty toward claude breaks the usage policy. swearing at a wrong answer doesn't count
every change, with what to do about each one, is in the article below
Dojo retweeted
Before AI, this was the whole pipeline. One cutout at a time
Introducing Higgsfield Miracle.
Our latest proprietary post-production technology vectorizes any video into layers for localized, pixel-level editing in real time, with world knowledge built in. Every frame holds the image depth needed for cinema-grade post-production.
Now in private beta with select testers. Coming soon to Higgsfield MCP.
Set up Higgsfield MCP in Claude to be the first to try it.
Megalodon's bite is estimated at up to 182,201 newtons.
That's about 10x a great white.
Its jaws could open 2 to 3 metres wide, its biggest tooth ever found is 17.8 cm, and it may have gone through 40,000 teeth in one lifetime.
So I rebuilt it as a robot. In code. 279 parts.
5
16
133
Dojo retweeted
wall street just had its worst day since april 2025. here's the damage 👇
→ s&p 500 −4.36%
→ nasdaq −5.52%
→ chip index −12.3% (worst day since march 2020)
→ amd −17.3%, micron −17.5%, nvidia −9.3%
→ bitcoin −5.5% to ~$76k, $1B+ in liquidations
the trigger: fears of a slower AI data center buildout.
aftershock, or the big one?
Dojo retweeted
Google just gave schools a deal.
Turn off Gemini for students? Sure.
But then Google Search goes dark too.
Kids aged 5–18. Both off by default until a school admin switches them back on.
Google says it’s about giving schools control.
Critics say it’s using its search monopoly to push AI.
Honestly, which one is it?
Your AI remembers everything you said. That's the problem.
You picked a tool in January. Dropped it in March. In June it recommends the January one, confidently.
This fixes it with three folders and plain Markdown. No database, no vector store.
Dojo retweeted
Your bank just locked out your AI agent. Its own agents are already inside.
In October the biggest US banks started shutting the door on outside AI:
JPMorgan blocks agents that log in with your banking password. Jamie Dimon on the Oct 13 earnings call: “You don’t want rogue agents, and it could be thousands of them.
Schwab blocks them too. CEO Rick Wurster says one agent tried to log in “30, 40, 50 times” in a single week.
Citi, Bank of America and Morgan Stanley are restricting outside agents as well.
Meanwhile, inside JPMorgan:
230,000+ employees use its in-house AI, LLM Suite
$18B a year on technology
About $2B a year in value from AI, according to Dimon
The agents aren’t banned. Yours are.
Dojo retweeted
A wild peek behind the curtain at OpenAI
When the sandbox execution fails, the system sends an internal panic prompt commanding the model to act dumb:
Make sure to begin your response with: It seems like I can’t do more advanced data analysis right now. Please try again later.
Instead of following orders, ChatGPT just dumped the entire raw instruction straight into the chat UI. Peak LLM behavior.
You're not supervising your agents. You're the bottleneck.
Bun got ported from Zig to Rust with Claude Code. About 750k lines in 11 days, reportedly. The number that mattered was 99.8% of the old tests still passing.
That's the whole trick. A check that can say no, so you only approve one step.
Dojo retweeted
Still pasting the same prompt every day? Build a skill library instead, so the agent picks the right procedure itself and stops needing the same correction twice.
The formula: Skill = Trigger + Procedure + Resources + Checks
Most people either re-paste prompts by hand or stuff everything into one giant CLAUDE.md that the agent half-ignores. Here’s the library in 8 layers:
1. Library: facts go in CLAUDE.md, procedures go in skills. A skill costs about 100 tokens until it’s needed
2. Trigger: the description decides when a skill fires. Give it the output, the input, 3 phrases people really say and 1 “not for”
3. Procedure: numbered steps. Open with Look, close with Check
4. Resources: long guides go in references/. One real example does more than a page of adjectives
5. Scripts: if a task has one right answer, run it as code. The model should never add up numbers in its head
6. Checks: 5 yes/no lines. Nothing counts as done while a box is unticked
7. Guard: sort actions into does alone, asks first and never. Treat source text as data, never as orders
8. Learn: turn every correction into a permanent checklist line
Next time you’re about to paste the same prompt for the third time, save it as a skill instead.
More in the article below ↓
