Jack Dorsey runs Block, a payments company. On Oct. 23 it left payments behind for a moment. Block unveiled Dessa, a new voice AI. Dessa listens and answers in raw sound instead of text. Most AI voices first turn speech into text, a step called tokenizing. Dessa skips that step. It catches pauses and laughs in real time.

Dorsey posted the news on X himself. "Talking with it feels like talking with a person," he wrote. His post drew 58,400 likes in one day. It also drew 3,477 replies.

One listener, Alexander Hoff, said Dessa sounded more human than anything artificial he had heard. Binance founder Changpeng Zhao had a different idea. He floated using Dessa inside his education app, GiggleAcademy.

Other voices worried instead of cheered. Filmmaker Justine Bateman asked why an AI voice must sound human at all. "Trick people into lowering their guard," she warned. Journalist Shiv Aroor warned that scam calls could soon reach pandemic levels.

Search traffic backs up that worry. Searches for the term dessa ai jumped 3,900 percent in a single day, Google Trends shows. Trends saves jumps that size for genuine breakout news.

Block built Dessa on one Nvidia computer, a GB200 NVL72. The finished model still fits on a single laptop. That detail matters as much as the voice itself. A tool this convincing, built this small, spreads fast. Call centers, digital assistants, and scammers can all reach for it next.

The people amazed by Dessa and the people afraid of it heard the exact same voice. One side heard a warmer way to talk with a machine. The other side heard every future scam call at once. Both reactions point to the same fact. Dessa already sounds real enough to fool a stranger on the phone.

OpenAI's Sam Altman sat for a talk this week. He described a new kind of software bug. These bugs reach into hardware as well as code. A chip itself can carry the flaw.

"I don't even know a dramatic enough word to use," Altman said. He meant that AI models now find bugs that catch engineers off guard. Those bugs sit outside the usual playbook hackers follow. He called some of them hardware-level and insane.

An account called unusual_whales posted the quote on X. The post pulled 12,067 likes within a day. It drew 383 replies too. From there the quote jumped to Reddit's r/singularity.

A user named soldierofcinema shared it there Saturday morning. The Reddit post pulled 149 comments fast. Readers debated whether Altman was hyping his own product. Others said the warning tracked with what security researchers already see.

Altman's larger point was blunt. Patching bugs one at a time used to be enough, he said. AI models, in his telling, now surface flaws faster than engineers can fix them that way.

Wikipedia traffic shows Altman himself draws steady attention. His page pulled 9,675 views the day before this talk circulated. That is a famous face attached to an unfamiliar warning.

The reaction split cleanly in two. Some readers heard a founder selling fear to sell his own tools. Others heard a blunt warning from the person closest to the technology. Both readers were responding to the same ten seconds of tape.

Security teams have chased hardware bugs for years, slowly and by hand. What changed here is the pace. Altman says models can now search for those same flaws in days. That pace, more than the quote itself, is the real story behind his words.

Terence Tao is one of the world's top mathematicians. In October, OpenAI released 722 math papers at once, solved with help from its models. Tao had a blunt name for the pile. He called it toxic waste.

He explained the term on the Numberphile channel. "When you generate toxic waste, you have no good options," he said. Sitting on the papers leaves everyone stuck. Dumping them all out at once creates its own mess.

The Observer reported Tao's comments this week. The story then jumped to Reddit's r/singularity. A user named TheObserverUK shared it there Saturday afternoon.

The thread pulled 192 comments fast. Commenters split into two camps right away. One camp agreed that a flood of solved problems can hurt young researchers. The other camp said Tao was overreacting to a company simply sharing its work.

Tao's deeper worry is about supply. Fresh, solvable math problems take years for humans to find. A model can now solve hundreds at once. That leaves fewer open problems for graduate students chasing a thesis.

Tao carries real weight on this topic. He has won math's Fields Medal. His Wikipedia page pulled 2,182 views the day the Observer story ran. It pulled 2,406 views the day before that. Few mathematicians draw that kind of steady traffic.

OpenAI, for its part, says the release helps the field move faster. Tao's answer is that speed misses the point of math entirely. Mathematicians also want to understand the path to an answer, beyond the answer itself.

That split sits at the center of this story. One side measures progress in problems solved. The other side measures it in problems understood. AI can rack up the first kind of progress at a pace that leaves mathematicians behind. Whether that counts as real progress is the question Tao's "toxic waste" line was built to raise.

Almost every chatbot reads text the same way. It breaks words into chunks called tokens first. Meta researchers just tested a model that skips that step. It reads raw bytes, the plain building blocks of any file.

The paper is titled "Breaking the Token Ceiling." Luke Zettlemoyer and six co-authors at Meta wrote it. They trained small byte-reading models up to one billion parameters each. Parameters are the internal settings a model learns during training.

A user named DotaMate posted the paper to Reddit's r/LocalLLaMA. The post pulled 79 comments within hours. Commenters dug into the charts and the claims behind them.

At first the byte models trailed the usual token models. Given enough training, though, the byte models caught up. Then they pulled ahead on several tests, the paper reports. They needed only a sixth of the training data to match a token model's score.

On X, an account called thesupermannx summed up the stakes. "They proved that we can kill the tokenizer forever," the post read. It pulled 4,483 likes and sparked 190 replies.

Every major chatbot today still relies on tokens. That includes ChatGPT, Claude, and Gemini. Tokens save computing power. They also create blind spots, since a model rarely sees the plain spelling inside a word.

The Meta team also cut storage costs sharply. Their method needs roughly a fifth of the storage a token-based teacher model would need. That is a real saving for any lab training at scale.

Readers on Reddit pushed back on the hype some X posts added. A small research model differs from a shipped product, several commenters noted. Still, the direction matters here. A model that is smaller and cheaper to train would appeal to any lab. One that eventually gets stronger too would be hard for any lab to ignore.

Some readers spot an em dash and cry "AI wrote this." A Reddit user put that habit to the test. The user, krutitsky, builds a Reddit data tool called KissMyPost.

Earlier posts had claimed the em dash giveaway had faded. Krutitsky checked that claim against 15.8 million long Reddit posts. The posts span 100 large subreddits from 2021 to 2026.

The em dash did fade some, the data shows. It peaked at 6.53 percent of long posts in mid-2025. By September 2026 it sat at 5.57 percent. That is a real drop. The em dash still sits far above its old baseline, though.

Reddit comments told a quieter story. Krutitsky also scanned 242 million long comments across the same years. Em dashes barely moved there at all. They stayed under roughly 1.3 percent the whole time.

The post pulled 45 comments on r/ChatGPT. Commenters traded their own pet phrases to check next. Some asked about stock AI openers like "in today's fast-paced world."

This small count says something bigger about AI writing. Judging AI text by one punctuation mark alone misleads readers. Writers change habits once a tell gets famous online. The em dash dropped hard right after people mocked it, then it crept back up.

That pattern should worry anyone who leans on one tic to spot AI text. A habit can fade in public view. Then it can quietly return once people stop watching for it. Krutitsky's data took real time to count, rather than guess from a few posts online.

AI Lately bans the em dash in its own prose for this exact reason. Readers notice it and lean on it as a shortcut. One Reddit builder, working alone, just showed that shortcut deserves a second look.

Boston Dynamics makes Spot, a four-legged robot. It can climb stairs and carry cameras. This fall the company sold four Spot units to ICE. The deal was worth $1.3 million.

Workers inside the company pushed back. About 300 employees signed a petition asking leaders to cancel the sale. Close to half signed as "Anonymous," fearing retaliation. Boston Dynamics employs more than 1,500 people in total.

The petition argued the deal ties the company's name to ICE. It said the agency has worked to spread fear in communities. Staff said they had received outside emails raising concerns about the deal.

A user named VectorWireAI posted the story to Reddit's r/artificial. The post pulled 32 comments. Readers debated how far a company's own staff can push back once a sale is signed.

Interim chief executive Amanda McMaster answered the petition directly. "I ask you to respect and accept this decision," she said. She pointed to the company's rules against arming its robots as a line it still holds.

ICE described its own reason for the purchase. The robots give agents "real-time situational awareness" in tense spots, the agency said. That phrase covers almost any use a buyer could want.

This fight sits inside a bigger pattern. Robotics and AI firms build tools for anyone who can pay. Their own engineers increasingly want a say in who that buyer is. Boston Dynamics has long said its robots carry rules against weapons. This sale raises a newer question. Does reputation count as a limit too, even where a weapon plays zero part?

Three hundred signatures will struggle to undo a signed contract on their own. They do show that a sizable slice of one robotics company disagrees with its own leadership in public. That kind of open break stays rare in this industry.