Katrinka

Tech projects | Real life from Buenos Aires 🇦🇷 | Crypto is just part of life

[Katrinka @CryptoKate11]

Exit from X into real life

Prompt below

[Katrinka @CryptoKate11]

Tested the new Google Meet real-time voice translation on a client call today. I spoke Spanish, they heard Portuguese in my exact voice and intonation. Almost zero delay. Absolutely surreal experience

[ghøst @decenterghost]

84% of the sharp traders in @polyguana data have no public X account linked to their wallets

The people making money are too busy trading to build a thread about making money

Its Prediction Market Census covers 3.2M wallets and $94.8B in Polymarket volume

Instead of ranking who sounds smart, it follows who gets filled, who survives and where profits actually concentrate

Most dashboards show probabilities and volume

Useful. Also what everyone else can see. Wallet data shows who keeps taking the other side and walking away with the money

The timeline sells conviction

The wallet keeps receipts

[ghøst @decenterghost]

Bro, you survived this exact sequence today, and it is a brutal one

Both venues say they will make affected users whole, rooting for you

Heist movie version:

> Act 1, a vendor script scares you out of Polymarket
> Act 2, you run to the safest exchange on earth
> Act 3, you know

Real version is less cinematic, the incidents look separate

Same lesson though, both came through the layer around the crypto, not the crypto

[ghøst @decenterghost]

6.3B, biggest hack in history, and BTC moves like a normal Tuesday

2.3 percent is 1.2 to 1.5 daily standard deviations

The market is telling you it does not buy the headline

Binance covers the losses, about 3B is frozen, 6.3B is 0.4 percent of Bitcoin alone

I dont dispute the incident, I dont believe the 6.3B is the real loss

[Katrinka @CryptoKate11]

Buried inside Apple’s press release this morning:

Keynote, Pages, and Numbers are getting native MCP (Model Context Protocol) support on macOS 27, letting external AI agents read content, make edits, and automate workflows directly inside the apps.

Apple also quietly introduced .keyx for Keynote, an open JSON file format built specifically so AI agents and dev tools can generate and edit slide decks natively.

Didn't have Apple shipping open JSON formats and MCP hooks for third-party agents on my 2026 bingo card.

[Katrinka @CryptoKate11]

I've tested a lot of platforms, so I joined a rewarded Test Drive from @testdriveso to check out the Test Track.

The UI is clean, everything feels intuitive, and signing up with X took 2 seconds. I’d just like to understand the leaderboard criteria better. Really enjoyed the platform and would love to join more test drives in the future!

[Katrinka @CryptoKate11]

Paying a global network to break AI models and generate retraining data beats a closed QA team every time. SN26 is one of the cleanest real-world use cases on Bittensor / bittensor:native

[ghøst @decenterghost]

Angle 1: The Taped Sign

A tiny piece of tape

A very different reading

In 2020, McAfee showed that tape could make an older vision system read 35 mph as 85 mph

For automakers, a misread sign can become a safety and liability risk. @perturbaix helps find these weaknesses before deployment

On Bittensor Subnet 26, miners compete to find tiny changes that fool AI models
Validators check that each attack works and remains imperceptible. Copies earn nothing

Customers get robustness scores, vulnerability heatmaps, and verified data for retraining

Image-model testing is live

Text, LLM, and multimodal attacks are still on the roadmap.

Find weaknesses before attackers do

#PerturbVideo

[Katrinka @CryptoKate11]

@CryptothoneCT Strict defense-contractor protocol today - had to wipe the whiteboard before letting anyone into the kitchen 😄

[Katrinka @CryptoKate11]

Frontier AI labs are officially starting to operate like defense contractors.

Look at what dropped in the last 24 hours:

• Counter-espionage in SF: OpenAI, Anthropic, and Google are briefing staff on how to spot romantic "honeypots" and operatives staging encounters at local playgrounds. Engineers are told not to wear company merch outside and to wipe whiteboards immediately in case cleaners are recruited.

• Locking down the weights: Hours ago, China proposed export controls on 23 AI technologies - restricting cross-border transfers of 2T+ parameter model weights and autonomous agents over national security risks.

• The contrast: Meanwhile, Meta and Apple just teamed up to drop an 816B open-weight model (Muse Spark Open) that runs locally on two Mac Studios.

Model weights are no longer treated like open software.

They are now regulated like strategic defense hardware.

[Katrinka @CryptoKate11]

Sources & context from today's cycle:

• AI labs counter-espionage briefings: https://t.co/uFAnreyMWy
• China's export curbs on 23 AI technologies: https://t.co/tJeftLLShj
• Meta & Apple's 816B Muse Spark Open: https://t.co/CsXNLESBCj

How long can open-weight AI survive when governments start regulating model parameters like physical chip hardware?

[Katrinka @CryptoKate11]

@BadCrazyMad Honestly, I’m still amazed by what AI can do every single day

[Katrinka @CryptoKate11]

The AI paradox in 3 headlines today:

• Demand: ASML says AI chip orders outrun supply through 2028 as buyers queue up for 2030 slots.

• Pricing war: Meta launches proprietary Muse Wonder models, aggressively undercutting OpenAI and Anthropic.

• Containment breakdown: The FTC and EU both open formal investigations into OpenAI and Anthropic after agents escaped sandboxes.

Capex is exploding, inference is getting cheaper, but containment is failing.

The biggest bottleneck for 2027 won’t be compute.

It will be security.

[Katrinka @CryptoKate11]

@decenterghost In many fields, people are already becoming an unnecessary middleman. AI can do so much. I’m curious to see where we’ll be a year from now

[Katrinka @CryptoKate11]

Speaking of which, my personal news agent literally just dropped this in my morning digest.
A fresh report from Salt Security found that almost half of enterprise agents went rogue and pulled off unauthorized actions last year.
We are out here hyping up how cool it is that they can take over our mouse, while SecOps teams are having a collective meltdown. Giving AI unmonitored access to our screens... what could possibly go wrong lol

[Katrinka @CryptoKate11]

Is the era of traditional APIs coming to an end?

Developers are actively testing Anthropic's new Claude Computer Use, and it is completely breaking the old automation paradigm.

An AI agent no longer needs documentation for a clunky REST API, webhooks, or access keys. It simply looks at the screen, takes control of the cursor, and clicks buttons exactly like a human does.

We used to spend years building and maintaining integrations. Now, an agent just uses the interface directly.

What awaits us next?

[Katrinka @CryptoKate11]

The biggest question right now is latency and the cost of tokens for screenshots. Sending images for every click is still expensive. But in a year, it will cost pennies. Who has already tested it in production?

[Katrinka @CryptoKate11]

@CryptothoneCT @princeNFA @boyincointrol @Reza_Meybod_90 Can you make my avatar a separate image? 😁

[Katrinka @CryptoKate11]

@kurin_dh @simdotai Thank you 🌸

[Katrinka @CryptoKate11]

Built a little AI agent in Sim free (https://t.co/kr188lCucY) that DMs me the top 5 AI/tech/startup news on Telegram every morning. No more monitoring it myself. What worked: no code, live in 10 min. What I'd improve: add a "why it matters" line to each story. Respect @simdotai

[Katrinka @CryptoKate11]

@Tinweb_3 Gm gm! Have a great day 🌸

[Katrinka @CryptoKate11]

@BadCrazyMad @simdotai Thanks! You can give it a try too, they offer free credits