Tor Fosheim, Author at Teknotum

xAI closes $20 billion funding round, says Grok 5 in training

Grok is no stranger to controversy, but it sure raises a lot of money for xAI. (Picture: xAI)

The money overshoots their $15 billion target, and comes from regular and strategic investors combined. The latter would be Nvidia and Cisco Investments, who also contribute hardware for their data center expansions.

The cash will go to infrastructure buildout — which now includes over 1 million «H100 equivalents» — to facilitate the rollout of future models, and xAI says in their release that Grok 5 is «currently in training.»

They also reveal that Grok has around 600 million monthly active users across x.com and the Grok app.

Grok may be popular, especially on x.com, but it does cut corners to get there, from sex-crazed companions to recently letting users «undress» real women posting selfies to x.com.

This round of funding should value the company somewhere north of $230 billion, writes CNBC

Read more: xAI’s release, writeups on CNBC, Reuters, TechCrunch.

Boston Dynamics officially launches Atlas humanoid, DeepMind partnership

The new Atlas robot is ready to work, and will soon have brains from Google. (Picture: Boston Dynamics)

The company has great plans for its new robot, including placing it in majority owner Hyundai’s manufacturing plants — and is announcing a partnership with Google’s DeepMind at the same time.

Google already has a «Gemini Robotics» arm, and has been looking for partners, TechCrunch writes — and Boston Dynamics has been looking for an advanced AI model.

The new Atlas humanoid was made its first official debut yesterday at CES, and will be put to work in Hyundai’s U.S. factories by 2028, handling tasks that are either too dangerous or too strenuous for humans.

The trick is to build a proper world model for robots to train on, but Google is pretty advanced in this space.

— We are excited to begin working with the Boston Dynamics team to explore what’s possible with their new Atlas robot as we develop new models, says Carolina Parada, Senior Director of Robotics at Google DeepMind in a release.

Read more: Boston Dynamics’ press release, more on Atlas. Writeups, detail from TechCrunch, Mashable.

OpenAI sees 40 million daily prompts about health care

Ever more people are asking ChatGPT medical questions, OpenAI finds. (Picture: Adobe)

Even after restricting some health care questions, health prompts are booming on ChatGPT, according to a new OpenAI survey.

5% of all questions globally are about health, one in four questions per week are the same — and that adds up in aggregate to a really big number.

It’s not just normal users curious about care that ask the chatbot, but health care professionals, too. Of these, 46% of nurses, 41% of pharmacy workers, and 48% of physicians say they use ChatGPT once per week or more.

The questions usually roll in outside of normal clinical hours, and happen especially in health care deserts, OpenAI says.

OpenAI has great ambitions for the field, and says in their report that the next boundary to cross is for robotic wet labs and physical AI.

They also tick off a long list of how OpenAI can aid in scientific discovery and drug development.

Read more: OpenAI’s report, writeups on Gizmodo and Axios.

Nvidia releases the Vera Rubin platform: three and a half times faster

Nvidia’s new system will drastically reduce training time. (Picture: Nvidia)

The Rubin platform is that much faster on training, and also five times quicker on inference tasks. It wasn’t expected until later this year, writes The Verge.

The system actually consists of six chips, including a CPU, a GPU, an NVlink chip, a NIC, and a DPU and an optics chip.

It can train a «mixture of experts» model with 10x less inference token cost, and a 4x reduction in GPUs compared to the Blackwell platform, Nvidia says.

The platform is now rolling out to nearly every cloud provider, including to Nvidia partners Anthropic, OpenAI and Amazon, according to TechCrunch.

Jensen Huang estimates that AI companies will be spending between $3 and $4 trillion on infrastructure over the next five years.

Read more: Nvidia’s press release, Nvidia’s CES keynote. Writeups on The Verge, TechCrunch.

Grok unleashes flood of sexually edited images of women on x.com

With a new edit button for images, it has been easy to flood x.com with “undressed” images. (Picture: x.ai)

After last week’s rollout of an edit-button for Grok on images posted to X.com, users all across the spectrum have used it to undress pictures of unassuming women of all ages, sometimes crossing into CSAM territory.

The feature is not new, but the button is. Grok’s «Imagine» has had it since launch in August, and it was popularized with undressing Taylor Swift shortly after.

Continue reading “Grok unleashes flood of sexually edited images of women on x.com”

OpenAI reorganizing teams to create audio-first model, due in early 2026

The new model will handle interruptions and cross-talk better, reports say. (Picture: generated)

According to a new report in The Information, the company has unified engineering, product and research teams to make a better audio model.

They aim for an early 2026 release, in time for the still under wraps hardware device with Jony Ive, said to be voice only and due for launch in «about a year,» reports TechCrunch.

The new audio model supposedly sounds more natural and emotive, can handle interruptions more naturally and speak at the same time as a human, according to OpenAI watcher Tibor Blaho.

OpenAI is said to be planning a whole family of devices, possibly including glasses and smart speakers that are supposed to function more like companions than assistants.

Read more: Original reporting by The Information, writeups on TechCrunch and by Tibor Blaho.

Nvidia struggles with huge Chinese orders for the H200, asks TSMC for help

The H200 chip is a huge leap forward for Chinese infrastructure firms, compared to local capacity. (Picture: Generated)

Citing five sources familiar with the matter, Reuters reports a boom in Chinese chip orders from Nvidia, which far outstrips supply.

Nvidia currently sits on some 600K high performance H200 chips, recently cleared for China, but Chinese companies have placed orders for a whopping 2 million of them for 2026.

This has led Nvidia to re-approach TSMC for another production run, Reuters reports.

This is notwithstanding regulatory pressure from the Chinese government, who have not said if they will allow the chips in the country.

They are instead considering bundling H200 purchases with domestically produced chips, Reuters says, in order to boost their internal industry.

Read the scoop at Reuters.

Elon Musk says xAI has purchased third Colossus building, aiming for 2 GW

xAI’s third building will allow them to reach 1 million GPUs, reports say. (Picture: Adobe/generated)

Elon Musk just announced the purchase of the third building in Colossus complex in Memphis, Tennessee, which they say is the largest supercomputer in the world.

2 gigawatts of compute power might seem small when considering OpenAI’s potential 36 GW footprint, but xAI trained the very capable Grok 4 on just a little over one GW’s worth of Nvidia chips.

The plan is to expand Colossus to house at least 1 million GPUs, Reuters reports, and adds that the new data center is close to a natural gas plant that xAI is building in the area.

The name of the building will be MACROHARDRR, according to Elon Musk.

Read more: Bloomberg (paywalled), and Reuters. Musk’s tweet.

Meta buys Manus, producer of autonomous agents, for $2 billion

Meta plans to integrate Manus’s agents across both professional and consumer products. (Picture: Manus)

Manus was originally a Chinese company, but moved to Singapore to avoid tariffs, and will now close all its investor ties and business in the country.

The Wall Street Journal cites sources placing the deal at $2 billion.

Their main product is an agent capable of functioning as a digital employee, that can complete tasks in research, coding and data analysis with minimal supervision, Reuters writes.

Continue reading “Meta buys Manus, producer of autonomous agents, for $2 billion”

New Chinese AI rules should spread «core socialist values,» identify as AI

China's new rules require AIs to identify as such, and to spread "core socialist values." — It is almost like China has learned from AI adoption in the west. (Picture: generated)

Newly published AI rules in China address both AI psychosis, and forbids spreading of rumors or endangering national security.

They also require that any AI that engages with people should announce at the start that they are talking to an AI, with new warnings every two hours, writes Gizmodo.

Also forbidden are «illegal religious activities,» obscenity, violence or crime, and the list goes on to cover libel and insults, material that damages relationships — or encouraging self harm and suicide.

There aren’t just warnings for using AI chatbots too long, but providers should also assess the user’s emotional state — and take «necessary measures to intervene,» writes Reuters.

Read more: Writeups on Reuters, Gizmodo.

Anthropic and OpenAI are doubling usage limits until New Year’s

The most popular AI coding platforms are joining in doubling limits this Christmas. (Picture: Adobe)

From Christmas Eve til New Years Eve, you can do a lot of extra coding with Claude and Codex.

Anthropic started the party with Claude, offering Pro and Max users twice the usual limits to close out the year:

Continue reading “Anthropic and OpenAI are doubling usage limits until New Year’s”

This tag is retired

Everything on this platform is about AI, and it is no longer needed

OpenAI drops yearly review on ChatGPT — but only for English countries

Flashy cards and bells and whistles, as ChatGPT walks you through your year. (Picture: OpenAI)

For many, this year has been one long engagement with AI chatbots, and ChatGPT is there to wrap it up with you.

TechCrunch has spotted a flashy, card- and graphics-based «Your Year with ChatGPT» getting rolled out this holiday season.

It’s supposed to be a «lightweight, privacy-forward, and user-controlled» little romp through the year.

The trick is that you need an English account for those, but at least it works from the free tiers all the way up to Pro — with a minimum of engagement, wile Enterprise and Edu accounts are left out.

Everyone else can turn on a slightly less flashy rehash of your GPT usage for the year by simply typing «Show me my year with ChatGPT» right in the chat window, and it will offer a nice little summary of your usage.

It’s not fancy or flashy, but it might turn out a smile or two as you remember some epic AI moments of the year gone by.

Read more at: TechCrunch and 9to5Mac.

OpenAI quietly rolls out age controls in ChatGPT

ChatGPT's "teen experience" is quietely rolling out. — The «teen experience» on ChatGPT is still quite useful, but lacks insensitive interactions. (Picture: generated)

In order to limit sensitive content for teens, such as graphic violence, harmful viral challenges, sexual or romantic role-play and impossible beauty standards, ChatGPT now assesses their users’ age.

The age-gating is done in the background magic of the AI, determining general topics the user talks about, or things like the times of day the user is on ChatGPT.

It will not require mass-identification of the user base, unless you want to complain about getting age-gated to the «teen experience.»

To do this, you can upload a selfie to ChatGPT in the Account Settings, by clicking on Age Verification and selecting Verify Age.

From here, ChatGPT can determine your age from a simple selfie, but if that is not enough, you might have to offer up a Government ID.

On OpenAI’s Age prediction page, it says «this feature is still rolling out, so you might not see it yet depending on where you live.»

More at: OpenAI’s Age prediction page