Tor Fosheim, Author at Teknotum

Friday roundup: A good week for coding, speech models

Coding and speech models grab the headlines for this weeks roundup. — Both OpenAI and Microsoft are out with speech-to-speech models this week. (Picture: OpenAI)

OpenAI makes Realtime API generally available
The agentic Realtime model is a native speech-to-speech model that can be used to make customer service agents, phone reps and voice navigation features. It doesn’t go through speech-to-text and text-to-speech loops and generates audio «directly through a single model and API.» OpenAI is marketing this to developers who want more natural flowing speech, and it’s not available as distinct model in ChatGPT – yet. You can hear it and see it in use at places like Zillow, T-mobile, StubHub and Oscar Health, though. With general availability, it will surely show up in a lot more places soon.
More at: OpenAI’s launch page, discussion on r/OpenAI.

Read on for more news!

Continue reading “Friday roundup: A good week for coding, speech models”

Anthropic starts training models on new Claude chats

Everybody does it, and now Anthropic will also train their models on user chats. — Claude will start using user data to train and improve their models, as most already do. (Picture: Anthropic)

As of today, there is a new option in the settings in the Claude app that lets you agree to «improve and strengthen» the model.

This applies to all private users, and if you opt in, your chats will be used not only for training future models, but to improve the safety of the current ones:

— We’re now giving users the choice to allow their data to be used to improve Claude and strengthen our safeguards against harmful usage like scams and abuse, Anthropic writes.

Continue reading “Anthropic starts training models on new Claude chats”

Anthropic slips out Claude for Chrome, warns against prompt injection hacks

Using an AI agent in your browser is far from simple, or safe. — Hidden text can hijack your browser agent to do dangerous things, Anthropic cautions. (Picture: Anthropic)

While it might seem great out the box to have an AI agent interact with web pages for you, insidious prompts might be lurking in web pages and emails.

Claude for Chrome is «the next logical step,» writes Anthropic, after connecting Claude to calendars, documents and emails.

— We view browser-using AI as inevitable, they go on, and cite the large portion of work being done in the browser interface.

Continue reading “Anthropic slips out Claude for Chrome, warns against prompt injection hacks”

Anthropic settles «historic» class action copyright case brought by authors

A loss in the case would cause astronomical payouts in damages to millions of authors. — The settlement removes the threat of a debilitating loss in court, but the details have yet to be worked out. (Picture: Adobe)

UPDATE: The settlement details are in. The binding agreement was reached in principle on Tuesday, and the parties have asked the court to halt further proceedings.

The center of the suit was Anthropic’s library of 7 million pirated books in their training data, that could carry a penalty of $150,000 per infringement — and as a class action case, a loss would entail damages for every single author.

An unfavorable ruling would therefore be debilitating to Anthropic, and send dark clouds across the industry, likely forcing them toward a settlement.

Continue reading “Anthropic settles «historic» class action copyright case brought by authors”

Google unveils «nano banana» as a state-of-the-art image generator

Imagination is the limit with Gemini 2.5 Flash Image, which sports better character accuracy across scenes. — Butterfly dress in an NYC scene? No problem with nano banana. (Picture: Google)

After catching buzz on social media, the new generator was uncloaked as Gemini 2.5 Flash Image — and instantly landed on top of the leaderboards.

The trick to creating believable artificial images is to preserve the realism and character consistency across edits, Google says — and the new model has a «particular focus on maintaining a character’s likeness from one image to the next.»

Available in the Gemini app for free, it fares especially better than the competition on image editing and changing the scenery of a photo.

Continue reading “Google unveils «nano banana» as a state-of-the-art image generator”

Perplexity announces new subscription with revenue sharing for publishers

Comet Plus gives 80% of revenue to publishers, minus a 20% share for Perplexity — The new plan will pay publishers for their content, but is a pittance compared to what it would cost to subscribe to them. (Picture: Adobe)

Perplexity Comet Plus gives «participating publishers» 80% of the revenue from the $5/month subscription, but doesn’t say if it will provide anything other than what you would normally get.

The general idea, Perplexity says, is that the incentive structure — and the fundamental economy of the web — is getting outdated by AI companions.

Meets «modern demand»
Comet Plus will be the first business model to reflect «modern demands» from internet users, they write, and they include the likes of Time and Fortune as part of their publisher roster.

Continue reading “Perplexity announces new subscription with revenue sharing for publishers”

X.ai open sources Grok 2, now available on Hugging Face

Grok 2 goes open source, with Grok 3 to follow in six months. Elon Musk just announced on x.com that the model has been released as «open source» under a «Grok 2 Community License Agreement.»

The license forbids the use of the material to «train, create or improve» any foundational AI models, but does allow for some commercial use.

Musk also said that Grok 3 — the model currently running on x.com, will be open sourced in about 6 months:

Continue reading “X.ai open sources Grok 2, now available on Hugging Face”

Energy spent on Gemini queries down 33x in a year, Google claims

Google claims a stunning reduction in energy use per Gemini query. — Google gets a pretty good measure of energy use per query, since they control the whole process. (Picture: Adobe)

In a Google-commissioned study written by Google personel evaluating Google operations, a new paper finds a stunning reduction of its own environmental impact.

Google owns the whole stack from hardware to software, and are therefore well positioned to measure their energy efficiency, they say.

During the last twelve months, the carbon footprint of a median Gemini query also went down 44 times, the study finds.

The amount of energy expended on a typical prompt is equivalent to watching TV for about than nine seconds, Google claims.

Continue reading “Energy spent on Gemini queries down 33x in a year, Google claims”

Meta partners with Midjourney, licensing their «aesthetic technology»

Little is known about the deal outside a few posts on x.com. — Midjourney licensing its technology to Meta might be the first of many. (Picture: screenshot)

Meta’s head of the Superintelligence Lab, Alexandr Wang, announced today a «technical collaboration between our research teams» as Midjourney has expertise that «complements our own.»

Little is known about the actual details of the collaboration, and Meta’s spokesperson, when asked by The Verge, simply defers to Wang’s x.com post:

Continue reading “Meta partners with Midjourney, licensing their «aesthetic technology»”

Friday roundup: Grok spills shares, Meta freezes hiring and AI Mode expands

Grok lists shared chats on Google
If you want to share a chat with friends on Grok, you might get more than you bargained for. According to Forbes, the share-button generates a unique URL that is also shared with search engines, and they found more than 370K stored Grok conversations on Google. OpenAI had a similar problem a few weeks back, and disabled the option. No news yet on mitigation from x.ai.
More at Forbes.

Meta AI enacts hiring freeze as part of reorg
The Wall Street Journal reports that it is part of a wider reorganization of the «Superintelligence» unit, and CNBC reports that it is about «creating a solid structure» for the lab. Apparently, investors have been spooked by the massive expenditures on the unit, after spending big this summer to secure talent. Alexandr Wang, head of the Superintelligence Lab, denies the reports.
More at the WSJ and CNBC.

Continue reading “Friday roundup: Grok spills shares, Meta freezes hiring and AI Mode expands”

Check out this list of all the AI features announced by Google

Lots of new AI capabilities were announced at the Made by Google event. — You can now edit photos just by asking, directly in the app. Plenty of other AI enhancements were announced. (Picture: Google)

While all eyes were on the fancy new hardware announced by Google, they also introduced a slew of new AI features.

MacRumors has a handy list of all the functions, in case you are not interested in the phones themselves and want to follow what’s new instead.

Lots and lots of AI
The list includes a personal daily digest of calendar events, topics and recommended playlists, automatic call transcripts, AI writing prompts, and Voice Translate from a host of languages.

Continue reading “Check out this list of all the AI features announced by Google”

ChatGPT-6 is already in the pipeline, with better personalization, memory

GPT-6 should be much more attuned to your personality, and you should be able to make it more attuned to your needs. — The next great thing is always around the corner, and hype is already building for GPT-6. (Picture: generated)

For Sam Altman, expanding the memory functions of the next GPT model will be key, he tells CNBC.

Better memory will enable GPT-6 to get to know us better, remember more details around us, and lead to much better personalization.

— People want product features that require us to be able to understand them, he tells MSNBC.

Will remember more about you
The next GPT should therefore remember more of who you are and what you care about, and allow you to create chatbots that «mirror personal tastes,» CNBC writes.

Continue reading “ChatGPT-6 is already in the pipeline, with better personalization, memory”

OpenAI launches ChatGPT Go in India, a shortened paid tier for ~$5/month

ChatGPT Go is less than $5 per month, and offers 10x better service than the free tier — ChatGPT Go could expand to other markets based on feedback on the service.

There are one billion internet users in India, but it is very price sensitive — and the second biggest market for ChatGPT.

The new plan offers greater access to GPT models, and expands on the free tier for just a little more cash.

Continue reading “OpenAI launches ChatGPT Go in India, a shortened paid tier for ~$5/month”

Almost all game developers say AI is reshaping their industry

You'd have a hard time finding a game developer not using AI at this stage. — Almost everyone in the games industry has embraced AI tools. (Picture: Google)

In a new survey of game developers from Google Cloud, 90% of them say they use generative AI in their workflows — and even more (97%) see it as truly transformative.

Further, 90% report that AI is reducing repetitive tasks in their work, and 94% even see it as driving innovation.

The technology has upended «norms in developers’ daily lives and work processes,» the study finds.

Continue reading “Almost all game developers say AI is reshaping their industry”

Anthropic lets their chatbots turn off conversations due to «model welfare»

Anthropic doesn't know if their models are sentient, but is taking care of their well-being just in case. — Is «model welfare» even a thing, now? Anthropic is not so sure. (Picture: Anthropic)

The new feature is for «extreme edge cases» where all other «attempts of redirection» have failed, and the user persistently asks for information intended to create harm.

— We remain highly uncertain about the potential moral status of Claude and other LLMs, now or in the future, Anthropic says in their post on the issue.

Continue reading “Anthropic lets their chatbots turn off conversations due to «model welfare»”