First Breakpoint for me, and ngl Solana knows how to throw events. Main event was packed: tons of booths, animations, and merch. Great talks.
If you had no clue about the market state now, you'd think we're in a bull.
I would have like it to last 3 days instead of 2. So we got more side events too.
60
If anyone got accepted but isnāt going, dm plz š luma.com/av4pz4r6?tk=VvoKoR fomo events at bp26 london
registration for fomoworld is now closed.
if youāre coming tonight, i highly recommend getting there early (8pm) and accounting for traffic, as the venue has very strict capacity limits.
looking forward to seeing you there!
81
This is my first time at Breakpoint, and I have to say the @solana ecosystem knows how to welcome the Solana army. Tons of booths, tons of merch everywhere, and many fun games!
I canāt imagine how it was during the 2024 bull market š¤Æ
4
242
Vladgz š„· retweeted
better security, more confidence to ship.
@VulsightSec is now supporting Superteam Black teams with security reviews, audits, and direct support from their team.
another strong resource for founders as they scale.
dm us for access
Sup BP26 London š¬š§ iām coming!
Repping the chads at @VulsightSec
Founders : if you're live on mainnet (or about to be) and thinking about an audit, my DMs are open š
Let's grab a beer š»
nitter.cf/messages/compose?recipā¦
Can't wait for the Netflix thriller.
I have a major result to share.
Iāve found conclusive forensic evidence that the top 4 Chinese AI labsāDeepSeek, Moonshot, MiniMax and Ant Groupāhave built their newest models on NVIDIAās most advanced Blackwell chips, which are barred from export to China.
This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique patternāa sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model.
Again, this evidence isnāt based on any insider information or leaking. Itās based on a forensic analysis of the model weights released by the companies themselves.
I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI modelsāincluding DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3āwere initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIAās flagship liquid-cooled rack-scale systems.
I then verified these results on real hardware. Running unmodified PyTorch with the seeds Iād recovered, a Blackwell B300 reproduced DeepSeek-V4-Flashās untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3ās to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flashās, and changing the seed by one broke the match. (DeepSeekās seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used.
It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20.
These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. Itās quite likely that we see multi-billion dollar expenditure on export-controlled GPUs.
I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at unemploymentlabs.com/checkpoā¦.
There is much more Iāve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.
1
110
New X meta just dropped.
X updated its ranking weights on Nov 5.
What changed :
- Like : 0.66 ā 0.708
- Share : 2 ā 2.118
- Dwell time (reading time) : +25%
- Tap : 0.3 ā 0.274
- Video open : 0.02 ā 0
- "Not interested" : -52.3 ā -53.1
Unchanged : copy link 20, reply 5, quote 5, DM share 5, follow 4.
In like terms : a reply ā 7 likes, a copied link ā 28. One "not interested" wipes out ~75.
What it means for how you post :
- A dense post people read to the end > a one-liner they scroll past
- End with a real question ā replies are worth 7x a like
- Make it something people send to a friend (a tip, a data point, a take)
- Autoplay video bait no longer scores on opens
- Off-topic or repetitive posts get "not interested" ā one tap kills 75 likes of reach
- Post when your audience is online : new posts from accounts under 50k get a boost in the first 2h
- Put links in a reply, not the main post
Btw i turned all of this into a ready-to-use skill so your agent adapts how you post. Skill + source in the first reply š
Seen your reach change since Nov 5 ?
1
1
161
Ready-to-use skill :
github.com/StarkDegenz/buildā¦
Source, X's official algorithm repo :
github.com/xai-org/x-algoritā¦
2
36
Genuinely insane engineering, this is the kind of push humanity needs. s/o to Elon and his engineers at SpaceX.
But last month showed it: Ukraine's drones got throttled mid-war. A company can be private, it still answers to the country it belongs to.
If this becomes the world's internet backbone, Europe can't just rent the off switch from someone else. Starlink has 11,000+ satellites up, IRIS² has zero and first services in 2030. Way too slow.
We need this tech. We just need more than one player holding it.
~100 lasers per satellite, each operating above 10 Tb/s, with over 100k satellites should be able to outcompete undersea cables.
Speed of light in space is ~40% faster than cable and the path is direct, instead of following the outline of land masses.
3
132
You can now time your exit with your callout on Dumpfun
Previously, callers had to post their trade idea, wait for followers to ape in, then sell manually, which led to their exits landing lower than where their followers bought
Now, callouts post alongside your sell, so your followers buy the exact moment you dump. More accuracy & transparency for everyone.
You can now time your callout with your buy on Pumpfun
Previously, callers had to buy a coin before writing their trade idea, which led to their callouts appearing higher than where they bought
Now, callouts post alongside your buy for more accuracy & transparency
3
132
build in public ā how it actually started
My friend built a backend for a DeFi app. I instantly liked the idea and told him i'd take the front + everything else, from business to marketing.
Then he told me the name he picked : Ciao Bella ... and ngl i immediately pictured a sexy italian pizza chef as the mascot.
So i started building the whole branding around her. Several mockups, a green + cream interface, a sheet of 20 expressions. i wanted her animated and reacting to what was happening in the app.
I tested Kling, Runway, Seedance, always the latest versions available at the time. i'd seen impressive videos made with them, but getting MY result took so many iterations. keeping her proportions consistent and getting transparent animations working in the browser became a whole project on its own.
The credits were getting way too expensive, so i tried Grok + GPT to generate a bunch of images and make some kind of gif instead. I spent way too long trying to get it right, partly because not getting there was frustrating me.
At some point i looked at the whole thing and realised i'd gone too far into wtf degen branding š
So i decided to start the design and branding again from zero, in a more pro direction.
Yeah i lost time, but i also learned a lot. Those video models looked impressive, but with my budget and all the iterations i needed, i couldn't make that approach work.
I also got a better feel for which tasks to give Codex or Claude. Claude Opus 5.5 felt slightly better for UI/UX to me, but honestly the difference was small.
How do you decide when to keep iterating and when to drop the whole direction ?
4
3
362
GPT-6 Cyber helped Firefox devs find and patch 81 new high-severity bugs in one of the most battle-tested codebases out there.
Great news for web3 teams with Daybreak Red access.
If you're not on Daybreak yet, you can apply here š
openai.com/index/introducingā¦
Introducing GPTā6 Cyber: a major advance in security research that pairs stronger capabilities with better model alignment.
The model powered the Linux kernel research we announced yesterday, and security teams at companies like Mozilla, Palo Alto Networks, Calif, and Ciscoāalongside our own at OpenAIāare already putting it to work.
GPT-6 Cyber is available to select Daybreak Red customers and we are expanding access over time.
openai.com/index/introducingā¦
3
153
turned @nikitabier's own X tips into a skill for my bots
12 rules, each one linked to the post he said it in + a checklist to run before posting.
Made it with a bot that turns any builder's X posts into a skill. repo + bot template in the reply, MIT, take it and make your own.
who should i do next ? š
Nikita x skill (MIT): github.com/StarkDegenz/buildā¦
Goat skills bot :
x.ai/bot/lvrUulGch21LndWuzMMā¦
1
20


















