OpenAI's GPT-5 has landed with a thud despite strong benchmark scores and praise from early testers.

Why it matters: A lot rides on every launch of a major new large language model, since training these programs is a massive endeavor that can require months or years and billions of dollars.

Driving the news: When OpenAI released GPT-5 last week, CEO Sam Altman promised the new model would give even free users of ChatGPT access to the equivalent of Ph.D.-level intelligence.

• But users quickly complained that the new model was struggling with basic tasks and lamented that they couldn't just stick with older models, such as GPT-4o.

Unhappy ChatGPTers took to social media, posting examples of GPT-5 making simple mistakes in math and geography and mocking the new model.

• Altman went into damage-control mode, acknowledging some early glitches, restoring the availability of earlier models and promising to increase access to the higher-level "reasoning" mode that allows GPT-5 to produce its best results.

Between the lines: There are several likely reasons for the underwhelming reaction to GPT-5.

1. GPT-5 isn't one model, but a collection of models, including one that answers very quickly and others that use "reasoning" — taking additional computing time to answer better. The non-reasoning model doesn't appear to be nearly as much of a leap as the reasoning part.
2. As Altman explained in a series of posts, early glitches in the model's rollout meant some queries weren't being properly routed to the reasoning model.
3. GPT-5 appears to shine brightest at coding — particularly at taking an idea and turning it into a website or app. That's not a use case that generates examples tailor-made to go viral the way previous OpenAI releases, like its recent improved image generator, did.

Zoom out: GPT-5 took a lot longer to arrive than OpenAI originally expected and promised. In the meantime the company's leaders — like their competitors — kept upping the ante on just how golden the AI age is going to be.

• The more they have promised the moon, the greater the public disappointment when a milestone release proves more down-to-earth.

What they're saying: In posts on X and in a Reddit AMA on Friday, Altman promised that users' complaints were being addressed.

• "The autoswitcher broke and was out of commission for a chunk of the day, and the result was GPT-5 seemed way dumber," Altman said on Friday. "Also, we are making some interventions to how the decision boundary works that should help you get the right model more often."
• Altman pledged to increase access to reasoning capabilities and to restore the option of using older models.
• OpenAI also plans to change ChatGPT's interface to make it clearer which model is being used in any given response.

Altman also acknowledged in a later post recent stories about people becoming overly attached to AI models and said the company has been studying this trend over the past year.

• "It feels different and stronger than the kinds of attachment people have had to previous kinds of technology," he said, adding that "if a user is in a mentally fragile state and prone to delusion, we do not want the AI to reinforce that."

Meanwhile, critics seized on the disappointments as vindication for their long-standing skepticism that generative AI is a precursor to greater-than-human intelligence.

• "My work here is truly done," longtime genAI critic Gary Marcus wrote on X. "Nobody with intellectual integrity can still believe that pure scaling will get us to AGI."

Yes, but: OpenAI's leaders argue that their scaling strategy is still reaping big dividends.

• "Our scaling laws still hold," the company's COO, Brad Lightcap, told Big Technology's Alex Kantrowitz.
• "Empirically, there's no reason to believe that there's any kind of diminishing return on pre-training. And on post-training" — the technique that supports models' newer "reasoning" capabilities — "we're really just starting to scratch the surface of that new paradigm."

Go deeper: I spoke with ABC News and NPR's "Here and Now" about GPT-5's bumpy rollout.

GitHub CEO Thomas Dohmke announced yesterday that he plans to step down, with Microsoft opting not to directly replace the position, according to memos shared first with Axios.

Why it matters: GitHub, which has operated largely independently since Microsoft acquired it in 2018, has become increasingly important to Microsoft's overall strategy to woo developers to Windows and Azure, as well as to its suite of AI tools.

Driving the news: Dohmke told staff in an email, seen by Axios, that he is leaving to pursue entrepreneurial endeavors.

• In a separate memo, Microsoft CoreAI head Jay Parikh outlined a new structure that will see GitHub leadership reporting to several Microsoft executives.
• Microsoft developer division head Julia Liuson will oversee GitHub's revenue, engineering and support.
• GitHub chief product officer Mario Rodriguez will report to Microsoft AI platform VP Asha Sharma.

The big picture: The world has changed a great deal since Microsoft acquired GitHub seven years ago. At the time, buying the code repository site was seen mostly as an embrace of the open source world Microsoft once shunned.

• However, with the rise of generative AI, GitHub has grown into a central place for developers to do their work.
• For Microsoft, meanwhile, GitHub has been at the leading edge of a broader effort to add AI-powered copilots throughout its product portfolio.

A new report zooms in on the gigantic amounts of energy needed specifically for training large AI models, as opposed to just aggregate estimates of training and use (known as inference).

Why it matters: The report yields a clearer picture of localized energy needs when hyperscalers build data center clusters that train exceptionally big "frontier" AI models.

• It's a joint project of the Electric Power Research Institute and Epoch AI.

State of play: The power required for the largest frontier training runs will surge this decade, potentially reaching 1-2 gigawatts by 2028 and a whopping 4-16 GW by 2030 (though that high end is considered unlikely).

Stunning stat: "This demand would be highly significant, with the high end for a single model approaching 1% of total U.S. power capacity," the report finds.

• "The energy demands of training cutting-edge AI models are doubling annually, soon rivaling the output of the largest nuclear power plants," said Jaime Sevilla, director of Epoch AI, in a statement.

The big picture: Total U.S. power needed to serve AI is estimated at 50GW in 2030, a huge uptick from 5GW today, the report finds in tallying inference and training.

• "Forecasts suggest AI could consume over 5% of U.S. generation capacity by 2030," it states.

Read the full analysis.

• Judges have been catching errors in the work of lawyers who use AI, but now they're trying to use AI tools themselves to speed up routine courtroom work. (MIT Technology Review)
• Elon Musk threatened to sue Apple for allegedly favoring ChatGPT in its App Store over competitors like his Grok. (Reuters)
• xAI made a version of Grok 4 available for free for all users, while the highest-end version remains exclusive to subscribers. (X)
• General Motors is weighing a renewed push on driverless cars, potentially including the rehiring of some former Cruise employees. (Bloomberg)

The Valkyries' new mascot "hatched" at Chase Center last night, immediately displaying some fierce dance moves. Making baskets took a little work, but Violet eventually found her shot.