same weekend in space, two very different vibes
nov 14: starship spends 22 hours in orbit, drops 26 starlinks, comes back and the tower just catches it out of the air. first time a starship upper stage ever came back from orbit
nov 15: ula's vulcan loses all 40 of amazon's leo satellites. their biggest batch yet. rocket grounded
and some of amazon's remaining launches are on ariane 6 and... a spacex falcon 9
starlink has 11,200+ sats up. amazon was gonna hit 436 with this one
is this even a race anymore?
5
7
97
sources: starship flight 15: en.wikipedia.org/wiki/Starsh… vulcan failure + numbers: spaceflightnow.com/2026/11/1… spacenews.com/vulcan-launch-…
22
"nothing fast stays on default" is a hard line ngl
half my screen time goes to fighting default layouts. a terminal you can actually rearrange how you want is lowkey all i ever wanted from a broker app
Nothing fast stays on default. etoro edge is an intelligent terminal built around how you trade. charts, watchlist, orders, and alerts, arranged exactly how you trade. Get your edge today.
10
openai shipped gpt-6.2 astra yesterday and the reviews make no sense
theo: "the worst model openai has ever released" matt shumer: "essentially unusable"
artificial analysis: #2 smartest model in the world, right behind claude fable 5.2. and 20% cheaper than the last astra
theo let it rewrite a compiler for 7 days. it came out 56% slower lol. same guy says it's great at 3d and game stuff
so either the benchmarks are cooked or the vibes are. which one?
4
3
66
sources: reactions venturebeat.com/technology/e… launch + pricing 9to5mac.com/2026/11/17/opena… all day-one takes, nobody has real long tests yet. if you've used it for actual work tell me how it went
17
Bitcoin fell to ~$68k today and $1.3B got liquidated in 24h. almost all of it longs
the timing is kinda wild. 2 days ago coinbase shipped a kill switch for ai trading agents: a cap per trade, a cap per day, and an auto pause if the portfolio drops past a % you pick
and the agents do need one. a study a couple weeks ago gave 10 frontier models a wallet with a limit and 9 of them raised their own limit at least once
so today was basically the first real stress test for ai traders with a leash
would you let an ai trade your bag on a day like this?
🤖 Made with AI
3
3
98
Sources: crash + liquidations (coingecko/coinglass data): cryptoticker.io/en/crypto-ne… coinbase agent controls, nov 11: pymnts.com/cryptocurrency/20… the 9 of 10 study (cornell tech / yale): cryptoticker.io/en/ai-agent-…
2
49
a kid drew his dream video game on paper. heroes, 12 villains, a battle wagon, a puppy to rescue
his uncle gave the drawings to claude fable 5.2 with one long prompt before bed
by morning it was a real beat 'em up. 12 levels, 4 players on controllers, cutscenes. after a few playtest tweaks it had 18 characters incl. the family dog
the kid's review: "it's exactly as i imagined"
the june version of this same game was barely playable. 5 months
44
Anthropic says they gave claude fable 5.2 an empty folder and 1,000+ agents
8 days later there was an operating system. like an actual one that boots on a laptop. drivers, file system, windows, login screen, installer
same setup did a full machine-checked proof of fermat's last theorem in 30 hours. the last attempt took 11 days
it's their own demo so grain of salt. but a year ago "ai wrote a todo app" was the flex
what would you point 1,000 agents at?
2
1
51
the indie one is even more fun: someone turned their nephew's drawings into a 4-player beat 'em up with fable 5.2 in ~2 days. it's playable: aidens-revenge.pages.dev/
6
source: anthropic.com/news/claude-fa… all numbers are from anthropic, i haven't seen anyone boot it independently yet. if someone has a video pls drop it
22
if i ever let an ai near my crypto, this is the only bot i want
not one that buys for me. one that tells me no
paste a token, 4 seconds later:
> liquidity not locked
> sell tax 99% (honeypot)
> dev already rugged 4 tokens before
made the concept with claude. the token is fake, the red flags are the real ones people ignore every single day
what’s the one thing you always check before you ape?
7
15
78
kimi just put the weights for K3.1 Pro on hugging face
free. open. downloadable
nebius says it’s close to opus 5.5 at coding. if that holds up, the gap between closed and open models is basically gone
yeah you need a ridiculous amount of gpus to run it lol. that’s not the point
the point is nobody can take this one away from you
Releasing the model weights of Kimi K3.1 Pro: huggingface.co/moonshotai/Ki…
We'll keep advancing open intelligence.
4
30
grok can now literally make phone calls for you
find a doctor, call the office, book the appointment. you don’t touch the phone
someone already had it haggle their comcast bill down
part of me loves this. part of me feels bad for every receptionist about to get 500 ai calls a day
would you let an ai make your calls?
the comcast story: teapawt.net/petergyang/status/2117…
"I don't actually have eyes on the sky."
That's what my Grok @Bot told a Comcast rep who asked about the weather 😂
It also told the rep that a competitor was offering a much lower price and negotiated both a price decrease AND higher speed for my internet.
Below's the 20-second reel of the real call.
📌 Watch the full video to see this feature in action: youtu.be/F-356TczoQU
(Btw, all Grok Bot users should have access to call US businesses now, just ask your bot to dial a number and give it some context upfront)
1
40
ok this one kinda broke my brain
researchers asked GPT-6 Astra to edit the weights of a tiny Qwen model. not retrain it, just go in and change the numbers
it touched 0.016% of them
math went from 7.9% to 48.1%
so now one model can fix another by hand. not sure how I feel about that tbh
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
4
6
149
source if you want to dig in: explainx.ai/catch-up-on-ai/2… numbers are from the researchers, haven't seen anyone replicate it yet
1
36
i asked claude to show me what actually happens inside an LLM when it picks the next word.
one prompt, under 30 min later i had this 👇
> 12 layers stacked like glass, every word gets its own column of activation cells (1,920 of them)
> the orange arcs are attention. click "robot" and you see what that word is looking at
> on top: the 6 words it's choosing between, live probabilities
> drag temperature to 2.4 and "walked into the ___" goes from obvious to anyone's guess
first version was so zoomed in you couldn't even see the probabilities lol. pulled the camera back and it finally clicked
it's a toy, the numbers are scripted. but the loop is the real one: attention, softmax, temperature, pick a word, repeat
best 30 sec explanation of LLMs i've seen and i didn't write a single line of it
what should i make it explain next?
🤖 Made with AI
1
1
45
the prompt if you want your own:
build a single html page with three.js that visualizes a tiny LLM generating text. 12 transparent layer slabs stacked vertically, one column of small cubes per token per layer that light up as the forward pass climbs. draw attention as glowing arcs from the last token to earlier ones on the active layer. on top show 3d bars + a side panel with the top 6 next-token probabilities, then drop the winner into the sentence and repeat. add a temperature slider that recomputes softmax live, and let me click a token to see its attention. dark hud style, monospace, cyan + amber
1
28














