Physical Intelligence is building general-purpose robotic foundation models through a loop of data collection, training, and real-world testing, TechCrunch reported on January 30th. In the company's San Francisco space, off-the-shelf robotic arms practice folding pants, turning shirts inside out, peeling vegetables, and using a test kitchen as training ground. The startup has raised more than $1 billion, spends heavily on compute, and has avoided giving investors a firm commercialization timeline while testing systems with partners in logistics, grocery, and food production.

The clumsy hand is the honest image of robotic intelligence. Software can appear brilliant in language because language tolerates polish, omission, and speed. A kitchen does not. A zucchini slips, fabric folds badly, hardware breaks, and a task that looks trivial becomes a lesson in force, grip, angle, friction, and recovery. Physical Intelligence is betting that general robotic skill will emerge from enough bodies, environments, and failures. The wager is expensive because physical knowledge resists theatrical demos. A model cannot merely describe peeling. It has to hold the vegetable, lose it, adjust pressure, collect the shaving, and try again while investors wait for a machine that can finally touch the world without pretending it understands it.

CISA's acting director, Madhu Gottumukkala, uploaded sensitive contracting documents marked for official use only to ChatGPT, TechCrunch reported on January 28th, citing Politico. The uploads reportedly triggered automated security warnings designed to prevent theft or accidental disclosure from federal networks. Gottumukkala had been granted an exception to use ChatGPT while other employees were prohibited, and Homeland Security officials examined whether the uploads harmed government security.

The leaking exception turns AI policy into a hierarchy of permissions. A public agency can forbid a tool to ordinary staff, warn against data leaving the network, and still allow the person at the top to carry internal files into a commercial model. The danger is not dramatic espionage. It is administrative confidence meeting a product built to absorb language. A contract document marked for internal use becomes prompt material, and the security system records the contradiction after the gesture has already occurred. The episode also weakens the theater of institutional caution. Rules look different when leaders treat them as temporary obstacles, and the chatbot becomes the place where that difference leaves a log.

Google expanded prompt-based editing in Google Photos to India, Australia, and Japan, TechCrunch reported on January 27th. The feature lets Android users ask for changes in plain language, from removing objects and reducing blur to restoring an old photo, changing a friend's pose, removing glasses, or opening a subject's eyes after a blink. Google says the edits run inside the app and that C2PA Content Credentials will indicate when an image was created or altered with AI.

The edited memory moves image manipulation from specialist software into ordinary speech. A family picture no longer needs a trained editor, a menu of tools, or visible labor. It needs a sentence. That makes repair easier, especially for damaged archives and bad snapshots that people still want to keep. It also changes the moral temperature of the album. The unwanted motorcycle, closed eyes, bad background, awkward pose, and inconvenient object become defects waiting for instruction. Metadata may record the intervention, but the eye meets the improved scene first. Memory has always been selective. Google Photos gives that selectivity a command box and makes correction feel as natural as asking.

San Diego Comic-Con and the Science Fiction and Fantasy Writers Association tightened their rules against generative AI, TechCrunch reported on January 25th. SFWA revised Nebula Award eligibility after backlash, barring works written wholly or partly by large language models and disqualifying work if LLMs were used at any point in creation. Comic-Con changed art show rules after artists objected, moving from a partial allowance for AI-generated art to a clear ban on material created wholly or partly by AI.

The rejected future carries special force inside science fiction. The genre spent more than a century imagining intelligent machines, artificial minds, synthetic artists, obedient tools, rebellious servants, and technical miracles. Its writers now confront a machine sold in the language of their own inheritance, trained on the cultural archive, and offered as a shortcut through the labor that made the archive valuable. The ban is imperfect because software companies keep embedding LLM features into ordinary tools. It still names a boundary. A convention hall and an awards ballot are declaring that speculative culture cannot be reduced to a prediction engine fed by previous dreams. The future may enter the story, but it does not get automatic authorship.

Sparkli, an AI-powered learning app founded by former Google employees, is trying to turn children's questions into interactive learning expeditions, TechCrunch reported on January 24th. The app generates audio, video, images, quizzes, games, and choose-as-you-go adventures around topics a child selects or asks about. The company says it has tested the product in more than 20 schools, built a teacher module, and added safety rules for sensitive subjects, while targeting children ages 5 to 12.

The captive lesson borrows the grammar of games to compete with games. Sparkli's founders describe curiosity, but the product also uses streaks, rewards, quest cards, avatars, daily topics, and generated media that can appear within minutes. Education has always needed attention. The new bargain is that attention must be won with the same machinery that trains children to return to apps. That can make a difficult idea more vivid than a worksheet or a wall of text. It can also make learning depend on a private engine that decides which path, image, voice, and reward will keep a child moving. The classroom receives an expedition, and the child receives a lesson shaped like a retention loop.

Blockit, a startup founded by former Sequoia partner Kais Khimji and John Han, raised $5 million to build AI agents that negotiate calendar scheduling directly with one another, TechCrunch reported on January 22nd. Instead of sending availability links, users can copy the agent on email or message it in Slack. The system learns preferences such as nonnegotiable meetings, movable appointments, skipped lunches, location choices, and even tonal cues in emails that might change a meeting's priority.

The negotiated hour turns time into a social database. A calendar once looked like a private grid of obligations, evasions, habits, and status. Blockit asks that grid to speak with other grids and settle conflicts before the owner returns to the thread. The promise is relief from small coordination labor. The cost is a formal model of preference, where lunch, hierarchy, urgency, politeness, and fatigue become instructions a system can apply. That shift may be useful because scheduling is miserable. It is also revealing because the agent must learn the secret grammar of a person's day. The meeting appears on the calendar as a clean block. Behind it, two delegated machines have already decided whose time could bend.

GPTZero scanned 4,841 papers accepted by NeurIPS and found 100 confirmed hallucinated citations across 51 papers, TechCrunch reported on January 21st. The number is small against the conference's total citation volume, and NeurIPS said incorrect references do not necessarily invalidate the research itself. The embarrassment sits elsewhere. The papers passed through a prestigious AI venue whose reviewers are asked to flag hallucinations, while researchers working in the field still allowed nonexistent references into the record.

The false footnote is a minor error with institutional teeth. Citations are the credit system of academic life. They route reputation, establish ancestry, justify claims, and help decide who receives jobs, grants, invitations, and authority. A fabricated reference does not need to destroy a paper to damage the room around it. It adds a phantom witness to an argument and asks reviewers to patrol thousands of tiny doors under impossible pressure. The irony is precise because the mistake appears among people who know the failure mode by name. The conference does not collapse because 51 papers contained bad references. It becomes harder to pretend that expertise alone can catch every fluent invention before it enters the archive.

BioticsAI received FDA clearance for software that helps detect fetal abnormalities in ultrasound images, TechCrunch reported on January 19th. The company, winner of TechCrunch Disrupt's 2023 Battlefield competition, says its computer vision system supports fetal ultrasound quality assessment, anatomical completeness, automated reporting, and integration into clinical workflows. Founder Robhy Bustami told TechCrunch that the harder task was proving reliability outside ideal cases, including across patient subgroups with higher risk.

The prenatal screen places machine vision inside one of medicine's most charged images. An ultrasound is already a strange document, part clinical evidence and part family apparition, read by specialists while parents search the blur for reassurance. Adding AI changes the authority around that blur. The system does not replace the scan; it judges whether the scan is complete enough, whether anatomy has been seen, and whether a possible abnormality deserves attention. That can help in clinics where skill, time, and access are uneven. It also makes prenatal care depend on a second reader whose confidence must survive bodies, demographics, equipment, and poor images. Before a child has a name, the image may already have passed through a regulatory file and a model trained to decide what the eye might miss.

Signal co-founder Moxie Marlinspike launched Confer, a privacy-conscious alternative to ChatGPT and Claude, TechCrunch reported on January 18th. The service is designed so that conversations cannot be used for training or advertising because the host does not have access to them. Confer uses passkey encryption, Trusted Execution Environments, remote attestation, and open-weight models to process prompts while keeping the chat hidden from the operator. Its free tier is limited, while unlimited access costs $35 a month.

The sealed confession treats privacy as an engineering cost rather than a slogan in a settings menu. A chatbot asks for material people rarely place in ordinary search boxes, including fear, money, illness, desire, family conflict, and professional doubt. Marlinspike's design answers that intimacy with encryption and constrained access, then passes the bill to the user. The price is higher than mainstream plans because the system refuses the cheaper bargain of retention, targeting, and model training. Privacy becomes visible as friction, limit, and monthly fee. A person can still talk to a machine, but the host is built to remain outside the room.