OpenAI finally learns to spell words.
Also Google thinks and spammers win email.

The Graphics Department Can Stand Down Now

OpenAI has announced a major upgrade to its image generation capabilities within the GPT-4o model, a release framed as an unprecedented leap forward. The real story, however, is that the system can now reliably spell text inside of the images it creates. We can all agree this is a low bar for a multibillion-dollar company, but here we are. The ability to render "perfect text" accurately within a visual medium was an ongoing struggle for previous models, making this latest feature the equivalent of a junior employee finally figuring out how to print a logo without the color shifting to neon green.

The other noteworthy feature is what OpenAI calls multi-turn image refinement, which allows users to edit and iterate on an image through conversational prompts. In essence, the AI will now tolerate a passive aggressive stream of revisions like "Make the sunset less sad" or "Please adjust the font to be more professional, you know, like the one we talked about last month." This capability has begun rolling out to both free and paid users, meaning the cost of an artist has not decreased; it has merely shifted from a salary line item to a monthly subscription.

Google DeepMind Debuts AI With Decision Paralysis

In a move seen as a direct counter-filing to OpenAI, Google DeepMind announced its Gemini 2.5 Pro model, which touts a new emphasis on "enhanced reasoning." The model is described as a "thinking model" capable of reasoning through its thoughts before providing a response. This essentially means the AI will now meticulously document its thought process before generating the same generic answer, much like a tenured engineer who has learned to always CC legal on every single email.

The real showstopper is the massive 1 million token context window, with a promised 2 million coming soon. This allows the AI to "comprehend vast datasets" including entire code repositories. Put simply, Google has built an artificial intelligence that can read every single document in your shared drive before you ask it to schedule a meeting. The model is currently available to Advanced users and is already leading on key industry benchmarks, mostly because it has an absurdly long attention span for corporate documentation.

The One Thing Spammers Are Good At Is Documentation

Legitimate IT administrators are reportedly facing an existential crisis as new data confirms that email spammers are, statistically speaking, significantly better at configuring email authentication protocols like SPF, DKIM, and DMARC than most companies. The discussion suggests that while legitimate businesses follow the basic instructions in the manual and get rejected by big-tech filters, spammers treat the standards as a high-stakes, competitive challenge to be fully optimized for cash flow.

The irony here is rich. The very people trying to adhere to best practices for non-malicious email are failing to get delivered because the high-volume, financially motivated scammers have become the unintended experts in deliverability. One comment noted that this is what happens when a generalist admin stops at the point where the documents say "success," while the specialists are incentivized to beat the "secret internal factors" used by the major providers. The entire system is now built for compliance by the lowest element.

Briefs

• Micro-Economy Win: A "Stoop Coffee" business idea, which involves selling simple coffee from a residential stoop, is wildly successful. The market has apparently decided that a neighbor with a kettle is more reliable than a 10-minute walk to a billion-dollar chain.
• Tesla Sales: European deliveries of Tesla vehicles are down by 43% while the overall EV market saw a 31% rise. Apparently, you cannot sell a self-driving car in a region that has perfectly functional public transport.
• AI Traffic Overload: Developers are blocking AI crawlers, forcing some to block entire countries just to keep their websites from being exclusively accessed by robots. It seems that the only way to get a human user is to assume they are operating from a region the machines have not colonized yet. A great defense tactic.