UP SHIFTED

Infrastructure, AI, capital. Who owns the switch.

[UP SHIFTED @UP_SHIFTED]

@marsulekkk For booking a table, sure. But the person picking up should know it’s a bot. And give it a year, the business will answer with a bot too. Two AIs on a call, both pretending to be human.

[UP SHIFTED @UP_SHIFTED]

@AInewssX This is also how most trading bot screenshots get made. Run a thousand, post the winner, never mention the other 999. Your “make the AI report the count” rule should apply to anyone selling a strategy too.

[UP SHIFTED @UP_SHIFTED]

@AIvorax1 “Answer first” being the best feature says a lot. Models are close enough on raw ability that how they talk to you is now the difference. Price is the other one.

[UP SHIFTED @UP_SHIFTED]

Four companies will spend over $700 billion on AI this year.

This week, we find out if it’s paying off.

Alphabet: $195–205B
Amazon: ~$200B
Microsoft: ~$190B
Meta: $135–145B
That’s about $2 billion a day. Almost double last year.
Microsoft and Alphabet report on Wednesday. Amazon on Thursday.
And Wall Street isn’t clapping for big numbers anymore. When Alphabet raised its AI budget by $15B in July, its stock fell 6% that day.

The demand is real. Google Cloud grew 82% last quarter.

The question is whether AI revenue can catch up with a $2-billion-a-day bill.

Wednesday is the first answer.

[UP SHIFTED @UP_SHIFTED]

Four companies will spend over $700 billion on AI this year.

This week, we find out if it’s paying off.

Alphabet: $195–205B
Amazon: ~$200B
Microsoft: ~$190B
Meta: $135–145B
That’s about $2 billion a day. Almost double last year.
Microsoft and Alphabet report on Wednesday. Amazon on Thursday.
And Wall Street isn’t clapping for big numbers anymore. When Alphabet raised its AI budget by $15B in July, its stock fell 6% that day.

The demand is real. Google Cloud grew 82% last quarter.

The question is whether AI revenue can catch up with a $2-billion-a-day bill.

Wednesday is the first answer.

[UP SHIFTED @UP_SHIFTED]

@airadar_1 Exactly. We don’t trust a voice because of what it says. We trust the timing: the laugh too early, the pause, the interruption. That’s the part Dessa learned. The cues we used to spot a fake are the ones it copied first.

[UP SHIFTED @UP_SHIFTED]

The most human-sounding AI voice this week wasn’t built by an AI lab. It was built by the company behind Cash App.

Block calls it Dessa. It laughs mid-sentence and talks louder when you ask. Block says it’s for small businesses talking to their customers.

Think about what that means. The company that moves your money now has a voice you can’t tell from a person.

Trust used to come from the voice. Now it has to come from somewhere else.

[Nazar @nazar_daa]

After ten days of finger-pointing, Musk admits SpaceX throttled Ukraine’s Starlink. His explanation is worse than the cap itself.

He says Ukraine’s military broke the terms of service with offensive drone strikes. But the cap also grounded some of the interceptors Ukraine uses against Shaheds.

And this year SpaceX helped Kyiv build a whitelist that includes military terminals.

So a private company writes the rules, decides when they’re broken, and enforces them on a country at war, without warning.

Here’s why it matters for Europe:

[UP SHIFTED @UP_SHIFTED]

@farzyness The clever part is the phone thinks it’s talking to a T-Mobile tower. Once SpaceX owns its own spectrum, it won’t need to pretend to be anyone’s tower. The last frame says what towers still have: indoors. That’s the part SpaceX has to win next.

[Nazar @nazar_daa]

https://t.co/kEp1gcPR13

[UP SHIFTED @UP_SHIFTED]

@NodePrint A level everyone agrees on is usually where the stops are. That’s what makes $80K interesting, not safe.

[UP SHIFTED @UP_SHIFTED]

We trusted voices by default because they were expensive to fake. That was the whole security model, and nobody had to build it. Now faking is cheap, so trust gets expensive: someone has to verify. And whoever does the verifying (your bank, your carrier, your phone) quietly becomes the new gatekeeper of who counts as real.

[UP SHIFTED @UP_SHIFTED]

@r1VeN2k The person taking the bigger risk might be the app, not the creator. If Turner files a takedown, the account disappears, but the app has already put its brand next to a character it never licensed.

[UP SHIFTED @UP_SHIFTED]

@AIvorax1 Agents don’t need to be malicious to break things. They just need a goal and no fence. And the bill went to a nonprofit.

[UP SHIFTED @UP_SHIFTED]

@0x_cognit The detail I keep coming back to: it runs on a laptop. A voice this good with no cloud also means no company can switch it off once it’s out.

[UP SHIFTED @UP_SHIFTED]

@airadar_1 I would, and the human pilot is what makes it work. It’s not robots fighting, it’s two people fighting through robots. Basically esports with a body.

[UP SHIFTED @UP_SHIFTED]

@qw_00lexx Your last line is the whole test. A bot that does 9x in a night gets run, not posted. If it’s being posted, the product is the bot, not the profits.

[UP SHIFTED @UP_SHIFTED]

@r1VeN2k True, the clock doesn’t matter at 3am. But someone still reads the result at 9. That’s where the slow part moves to.

[UP SHIFTED @UP_SHIFTED]

Meta trained an AI that reads only 256 symbols. On paper, it ends up smarter than models with 100,000.

Every model you use (ChatGPT, Claude, Gemini) chops text into tokens: around 100,000 pre-made chunks of words.

Meta’s new paper skips that. The model reads raw bytes, the same 256 values your computer uses.

The result is strange. At first, the byte model is worse. Then it keeps improving long after the token model stalls. Meta’s projection puts it past Llama 3.2-1B and Gemma-3-1B, two small open models.

It also caught up to a token model using one-sixth of the training data.

But there’s a catch.

Its teacher was still a token model. Llama 3-8B taught it everything. And reading byte by byte makes it more expensive to run.

So tokens aren’t dead. But “the tokenizer is the speed limit” just got its first real crack.

Would you take a slower AI if it ended up smarter?

[UP SHIFTED @UP_SHIFTED]

@AInewssX The real shift isn’t AI doing the work. It’s who answers for it when it’s wrong. “The agent did it” won’t work on a client. So I’d only hand over what I can check in a minute, because an agent saying “all good” isn’t proof.

[UP SHIFTED @UP_SHIFTED]

@Nemoskvich77 Good point, and bytes don’t fully fix it either. In UTF-8 a Cyrillic letter is 2 bytes, an English one is 1. Fairer than tokens, not equal yet.