Teknotum - Page 3 of 44 -

More frontier labs agree to submit pre-release models to US Commerce Dept.

Google, Microsoft and xAI will be joining Anthropic and OpenAI in submitting their early models for testing by the Center for AI Standards and Innovation (CAISI).

— CAISI will conduct pre-deployment evaluations and targeted research to better assess frontier AI capabilities and advance the state of AI security, the Center says in a release.

This program was started under the Biden administration in 2024, and AI labs «frequently provide CAISI with models that have reduced or removed safeguards,» the National Institute of Standards and Technology says.

It was reported yesterday that the White House was considering some kind of vetting of early models before release to the public sector, while giving the Pentagon a first shot at «useful» cyber capabilities that might arise.

— Independent, rigorous measurement science is essential to understanding frontier AI and its national security implications, says CAISI Director Chris Fall.

Read more: NIST’s release, Reuters and Axios.

OpenAI releases GPT-5.5 Instant, with more concise answers

No more five-page essays when you need a simple answer. (Picture: generated)

Replacing GPT-5.3 Instant as the «daily driver» for hundreds of millions of users, the new model touts major changes to common gripes.

For those fed up with three-page essay responses to simple questions, the new model should offer «clearer, more concise answers,» with «stronger» and «tighter» responses that are more «to-the-point.» This should reduce the need for frequent follow-up questions, OpenAI says.

OpenAI says the new Instant will be «more effective» at leveraging history from previous chats, files and Gmail. It should also be showing you its sources when it uses memories.

Hallucinations are supposedly also way down, reducing the rate by 52.5% in internal tests, while inaccurate responses dropped by 37.3%.

The model will become the default option for most ChatGPT users on the Plus and Pro tiers, and is rolling out «over the next two days.» Other plans should get it «soon.»

The White House reportedly discussing vetting AI models ahead of release

The White House says any Executive Order will come from the President himself. (Picture: Adobe)

The Trump administration has appartently been spooked by the cyber capabilities of Anthropic’s Mythos model and OpenAI’s GPT-5.5 — and is considering an Executive Order to vet new models ahead of release, Axios and the NYT reports.

These models have both been limited for their ability in cybersecurity, and point to a not-so-distant future where such capabilities might be widely available.

To that end, the White House’s Office of the National Cyber Director held all of two meetings last week, with tech and cyber companies on the one hand and with trade groups in tech on the other, according to Axios.

The ONCD has also been discussing safety testing for federal AI deployments, by assessing the security exposure of AI models before rolling out to the public sector.

The NYT reported on this first, and is saying that there might be a safety review for new models, while giving the Pentagon the first shot at eventual «useful» cyber capabilities, but would not block their release.

Any discussion on «potential executive orders is speculation,» a White House official told Axios.

Read more: Axios and the New York Times.

OpenAI co-founder and President Greg Brockman reveals $30 billion stake

ChatGPT has been a lucrative endeavor for those at the top. (Picture: Shutterstock)

«The compensation was certainly secondary to the mission,» Brockman said in court yesterday, according to NBC News.

He was being questioned at the Musk trial by lawyer Steven Molo, asking «You just happen to be $30 billion richer?»

The OpenAI President also holds a stake in Sam Altman’s family fund and chip company Cerebras, which is a significant supplier to OpenAI, Reuters reports.

Elon Musk is suing OpenAI in a California court, arguing the startup he co-founded has abandoned its altruistic non-profit mission and going on an «enrichment spree» on a for-profit basis. He seeks to paint the executives as greedy.

OpenAI counters that Musk is driven by resentment of their success and an obsession to control the company. The trial is ongoing.

Read more: Trial notes by NBC News, Reuters, Business Insider, and Wired.

OpenAI debuts «pets» for Codex — and they are actually useful

Codex is pretty capable in itself, but sometimes you wish you could get status messages while it is working in the background, and maybe get it with a cute little mascot.

This «problem» has now been solved by OpenAI, by letting you use customizable «pets,» or mascots that float on top of your desktop and gives you useful status updates from Codex at work.

They can show you active threads, tell you whether Codex is running, waiting for input, or if anything needs your attention — at a glance, from different apps.

They can also be quite cute and the possibilities are basically endless.

It is even possible to make the little Finder guy, or download it from Petdex. There are also avatars for Dario Amodei, Sam Altman, Goku, and a goblin and Clippy on Codex-pets.org.

The Oscars move to protect human authorship in new rules targeting AI use

Even if using AI, there must be human creative authorship. (Picture: Oscars.org)

The Academy of Motion Picture Arts and Sciences just posted their rulebook for the 2026 Oscars and in it, they state that all movies under consideration must be human authored — while allowing some creative AI use.

Eligibility for the actors awards is that they are officially listed on the credits of the movie and that the role is «demonstrably performed by humans,» they unequivocally say.

The screenwriting awards also establishes that works must be credited and «human authored,» so no AI allowed there, either.

Continue reading “The Oscars move to protect human authorship in new rules targeting AI use”

GPT-5.5 found on par with Mythos, with OpenAI to limit access to Cyber version

It took about three weeks for a competing model to hit parity with Mythos. (Picture: Adobe)

After a major research paper by the UK’s AI Security Institute found GPT-5.5 a little better than Mythos, Sam Altman moved to limit access to the Cyber version of the model.

The paper probes «vulnerability research and exploitation against realistic targets and modern mitigations» through rigorous tests, and found GPT-5.5 had a pass rate of 71.4%, compared to Mythos’ 68.6% on the most advanced evaluations.

According to the AISI, their test suite proves that Mythos is not a one-off act of brilliance, but part of a wider trend for frontier models. They say «we should expect further increases in cyber capability from models in the near future, potentially in quick succession.»

At the same time, Sam Altman posted on x.com that OpenAI will indeed follow Anthropic’s lead on limiting access to GPT-5.5-Cyber to «critical cyber defenders:»

— We will work with the entire ecosystem and the government to figure out trusted access for cyber; we want to rapidly help secure companies/infrastructure, Altman wrote.

Read more: The AI Security Institute, Altman’s x post, TechCrunch. Discussion on r/Singularity.

OpenAI confirms and explains GPT’s affinity for mentioning goblins

It’s true; ChatGPT is overly attached to goblins and gremlins. (Picture: Adobe)

Two days ago, OpenAI caught a mini-viral moment after a system prompt was found to contain a total ban on the mention of goblins, leading to more questions than answers.

Today, OpenAI is revealing their research on the issue, and can reveal that this was indeed real. Starting with GPT-5.1, the models did definitively prefer using «goblins» in their replies.

The culprit was the «nerdy» personality, which debuted with the launch of the 5.1-model and had increased «goblin»-mentions by 175% and «gremlin» by 52%. And by GPT-5.4, «goblin»-use had balloned by 3,881.4%, causing consternation at OpenAI.

The error seems to stem from rewarding a «playful style» with creature references, and this has since propagated through later releases.

The «nerdy» personality was retired in March after GPT-5.4 was released, but goblins snuck into the training data for GPT-5.5, too — forcing the system prompt to «Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user’s query.»

Read more: OpenAI on goblins. On the system prompt: Gizmodo, Business Insider, and Ars Technica. Discussion on r/ChatGPT.

Meet Talkie-13B — a «vintage» language model with a 1930 knowledge cutoff

Talkie is not outright dismissing talk of splitting the atom. (Picture: screenshot)

While it can be fascinating to talk to an LLM «from the past,» it also prompts questions of scientific discovery and the limitations of large language models.

For example, you can ask it if Hitler will ever be Chancellor of Germany, or culturally, what it thinks of black folk.

The real question is perhaps whether it will be able to discover anything «new,» that we can test, as in what it thinks about splitting the atom, and if it will ever be possible. Will Talkie ever predict the new science — or the wars — of the 1940s?

The model itself was painstakingly made on collected data before its 1930 cutoff (because copyright expired), and only uses English data — from books, periodicals, newspapers, scientific journals, patents and court documents.

The result is an LLM that lives in the past, that might do scientific discovery and that serves at least as a nice cultural window on times gone by.

You can try it on Hugging Face, and download it from Github.

Read more: Introducing Talkie, Gizmodo. Discussion on r/Singularity.

The European Union starts process to open up Android to AI competitors

Gemini is basically enjoying a monopoly for integrated system access on Android. The EU wants to change that. (Picture: generated)

Google has been aggressively implementing AI and Gemini on its platforms, such as web search — and Android. Right now, Gemini is basically the only AI with system access on the platform, and the EU sees room for improvement.

Under the Digital Markets Act, Google isn’t just another vendor — it’s one of seven dominant platforms, deemed «gatekeepers» to other services. That means it has to behave like a platform, like Windows, and offer equal access to its services.

The European Commission lists letting competing AI assistants have easy access to functions like sending emails, sharing and editing photos — and have system level access to control apps. It should also provide Android API access and support for free, they say.

Continue reading “The European Union starts process to open up Android to AI competitors”

Google signs on to Pentagon contract for classified work

All top labs except Anthropic are now available to the Department of War. (Picture: Shutterstock)

Joining OpenAI and xAI, Google’s contract follows the controversial «all lawful government use» rules for classified work at the DoD, The Information (paywalled) was first to report.

The contract is quoted by Reuters to include specific language against use for «domestic mass surveillance» and autonomous weapons «without appropriate human oversight and control,» but Google doesn’t get to second guess or outright ban this usage of their models, The Verge reports.

Anthropic famously defected from its Pentagon contract over precisely these issues, but OpenAI quickly signed on to a contract with better language.

Rumors of a coming deal sparked protest internally at Google, with 600 Googlers signing a letter demanding that Google refuse any classified use, as they also did over Israel contracts.

Read more: The Information (paywalled), Reuters, The Verge, and TechCrunch.

Taylor Swift moves to trademark audio clips and image to protect from AI

The likeness of this image could be trademarked and used to sue AI platforms. (Picture: U.S. Patent & Trademark Office)

The trademark filings of two audio snippets and an image of her on stage with a guitar are intended to ward off AI deepfakes.

The problem that copyright doesn’t fix is not the actual copying of Swift’s work, but deepfakes that copy her voice and likeness in generated content:

— Historically, singers relied on copyright law to protect their recorded music, writes trademark lawyer Josh Gerben. — But AI technologies now allow users to generate entirely new content that mimics an artist’s voice without copying an existing recording, creating a gap that trademarks may help fill.

Trademarking an artist’s voice and image has only been done once before, by Matthew McConaughey, in an effort to ensure that any «use of his likeness would require approval» in an AI world, Gerben writes.

It’s a novel approach that has yet to be tested by courts, but should provide protection for similar copies of Swift’s works — even those that are «confusingly similar.»

Read more: Josh Gerben’s blog post, Variety, and Reuters.

OpenAI and Microsoft drop exclusivity

OpenAI’s models can now be distributed through AWS and Google Cloud. (Picture: Microsoft)

As Microsoft made big early bets on OpenAI, they gained exclusivity to their models and a revenue share from OpenAI.

Now, with a $50 billion deal from Amazon waiting in the wings, the companies are amending the deal so the models can be made available «across any cloud provider.»

That’s big news for enterprise adoption and competition with Anthropic, as their models can now be made widely available on Google Cloud and AWS.

As part of the new deal, Microsoft is still OpenAI’s «primary cloud partner,» and will get dibs on new models. The difference is the non-exclusive license through 2032, and that they will no longer pay a revenue share to OpenAI. OpenAI, however, will be paying Microsoft through 2030, with an undisclosed total revenue cap.

— With this, builders will have even more choice to pick the right model for the right job, Amazon CEO Andy Jassy wrote, according to Reuters.

Read more: OpenAI’s announcement, Reuters, and Ars Technica

OpenAI is building a «fundamentally changing» agent-first phone, Kuo says

This is how Ming-Chi Kuo imagines a new agent-first interface. (Picture: Ming-Chi Kuo)

Supply chain analyst Ming-Chi Kuo, mostly known for breaking early Apple news, has published an article on x.com claiming that OpenAI is already working with MediaTek and Qualcomm to develop a phone with an agentic interface.

— Smartphones will remain the largest-scale device category for the foreseeable future, he writes, and shipments of high-end phones are around 300-400 million units a year — a mass market OpenAI would be keen on reaching.

The phone itself should be an agent-first experience, «fundamentally changing how people think of smartphones,» Kuo says.

This means moving away from apps to agents, from icons to tasks, and then doing away with the well-worn grid interface in favor of an agent-powered stream layout.

Along with the chip giants on board, systems integrator Luxshare is slated as their main manufacturing partner, aiming for a full system spec late this year or early 2027. Mass production is «expected» in 2028, meaning the project is moving fast.

Read the full story on x.com.

Google to invest up to $40 billion in rival/partner Anthropic

Google and Anthropic are rivals in the chatbot arena, but partners in compute. (Picture: Anthropic)

Alphabet, Google’s parent company, has reached an agreement to initially invest $10 billion in Anthropic, increasing to $40 billion on reaching «performance targets,» Reuters reports.

The deal comes hot on the heels of Amazon’s investment of $25 billion this week, also contingent on «commercial milestones.»

Google’s investment comes as Anthropic is seeking to expand its compute capacity after several reports of throttling and downtime for its Claude service, amid rising popularity.

Earlier this month, Google and Broadcom secured a «partnership» with significant compute for Anthropic, as the company announced it had hit a $30 billion annualized run-rate revenue.

Read more: Reuters, CNBC, and TechCrunch.