Dr. Ryan Ries here. This week I have a ton of AI news stories to share. I’m never surprised, but it always is wild to me how quickly things change in this industry.
Today we’re talking about:
- Telepathy in the form of a wearable headband by 2027
- AI is being used to edit biology and it’s pretty amazing
- Tokenomics & Google’s newest model
- ChatGPT can track your work and make you more productive
- 2 new AI transparency laws
A researcher quit OpenAI to build telepathy
Naomi Bashkansky spent about a year and a half at OpenAI working on model safety and alignment. Then, on July 23rd, she left. The next day she started at a startup called Conduit as a founding researcher. Her job title might as well be telepathist.
Conduit is building thought-to-text models trained on noninvasive neural data. No implants and no surgery necessary. Just a headset that reads electrical signals through the scalp and a language model that fills in the gaps. Bashkansky's own prediction, stated plainly in her resignation essay: in a couple of years, the main way we talk to our AIs will be with our thoughts.
Here is her rough timeline:
By 2027 – A headband that turns your reactions into prompts for a coding agent.
By 2030 - Direct neural representations feeding AI systems.
By 2035 - Something that feels like a sixth sense.
This is really interesting to me because years ago, when I worked for Daqri, we acquired a startup called Melon that had created a headband to try and move a mouse or use your mind to play some sensory games. I worked on the hardware to try and improve the system and make it a reality. Let me just say -- it is incredibly difficult. That was 10 years ago, and tech has definitely changed a lot, so we will see if they had more luck than we did. It seems like they are not the only company trying to move this forward now.
A startup called Hemispheric just emerged with $52 million, co-founded by one of the engineers who helped build Apple's Face ID. Their model, called Descartes, has six billion parameters and was trained on 250,000 hours of brain data from over 100,000 people. The target is diagnosing depression, PTSD, and Alzheimer's from a 15-minute headset session, something medicine has never had an objective test for.
While the objectives of Conduit and Hemispheric are totally different, both bet that a frontier language model, layered on top of noisy brain signals, can finally make sense of the noise the way it made sense of messy internet text.
AI is starting to edit biology
Researchers at Stanford and the Arc Institute used a genome-generating AI model called Evo, sometimes described as ChatGPT for DNA, to design 302 new versions of a simple bacteriophage. This is not to be confused with macrophages that the human body also uses to collect the trash in your body.
(For those of you that are not in life sciences - Google says a bacteriophage (or phage) is a tiny virus that targets, infects, and reproduces inside bacteria).
Sixteen of them worked. They infected and killed bacteria in the lab, and some outperformed the natural virus they were modeled on. A few were different enough from anything known to science that they'd count as new species.
That is mind-blowing to me!
Antibiotic-resistant bacteria are a growing problem, and phages that hunt specific strains could become an actual treatment path.
Meanwhile, in a much cozier corner of biotech: a company called Kindred Companion Sciences used CRISPR to knock out the gene behind Can f 1, the main protein that makes people allergic to dogs.
(If you have not read or seen some of the cool movies that reference gene editing - Google says CRISPR stands for "clustered regularly interspaced short palindromic repeats". It is a powerful technology that scientists use to change the DNA of living things. It acts like a set of molecular scissors to cut, fix, or replace specific genes. I highly recommend Change Agent by Daniel Suarez if you are interested in fiction and what the future could hold with CRISPR technology).
Two beagles, Alfie and Bailey, don't produce it in their saliva, hair, or dander. Their creator, a geneticist who's been allergic to dogs his whole life, tested himself against samples from the edited dogs and from breeds long marketed as hypoallergenic, like poodles and goldendoodles. He reacted to all of them except Alfie and Bailey.
It's not the same field as AI-designed viruses, but it's the same underlying idea. We can now edit biology with a precision that used to take generations of breeding.
For me personally, my family is allergic to dogs so I found this pretty interesting!
Google just made its cheap model a lot more useful
Gemini 3.7 Flash landed on August 13th, just three weeks after Gemini 3.6 Flash. Google is calling it the most intelligent workhorse model yet for coding and agents, and the benchmark gains back that up: a jump from 34.4% to 43.6% on FrontierCode, and from 17.0% to 30.4% on AutomationBench, which measures real business workflow automation.
For folks thinking about tokenomics, Google cut the price in half: $0.75 per million input tokens and $3.75 per million output tokens, down from $1.50 and $7.50.
But I caution you to read the fine print.
That's only an introductory rate through the end of 2026. On January 1st, 2027, it doubles right back to where it started. So plan your 2027 costs around the standard rate, not the launch day headline.
ChatGPT wants to watch you work
OpenAI rolled out a feature called Computer History for the ChatGPT desktop app on Mac. It's opt-in, and it replaces an earlier preview called Chronicle that worked by taking screenshots. Computer History skips the screenshots. Instead, it logs clicks, keystrokes, app switches, and browsing activity through macOS's accessibility system, then turns that into a searchable timeline and local memory that ChatGPT and Codex can reference.
While this kind of sounds creepy, it does have some benefits for your productivity.
You can ask what you were doing before your last meeting, and ChatGPT can point you back to the file. If it notices a repeated task, it can offer to automate it. OpenAI says it doesn't capture your screen, your microphone, or anything from private browsing tabs.
Again, I am throwing up a yellow flag here.
OpenAI's own documentation says the memory files sitting on your Mac aren't encrypted, and other programs running under your account could potentially read them. The company also flagged its own increased risk of prompt injection, since instructions hidden on a website you visit could end up feeding into what ChatGPT sees.
Opt-in features that log this much of your day deserve the same scrutiny you'd give any new piece of software touching sensitive files. If you’re interested in trying, I’d turn it on for a couple of low-stakes apps first and see what it does.
Latest in AI Transparency Laws
The EU AI Act and California's AI Transparency Act both took effect recently, requiring advertisers to disclose when AI generated a depiction of a person, place, or object. Experts are comparing the moment to GDPR: legally real, with the practical rollout still getting sorted.
For this one, check on your marketing teams, they might be crying:
Anthropic is now weaving invisible watermarks into text generated by Claude models released on or after August 2nd. The mark doesn't change the meaning or readability of the text, and it travels with the writing when it's copied and pasted, though a heavy rewrite can knock it loose. It's part of Anthropic meeting its commitments under the EU's Code of Practice on AI transparency, and it's rolling out worldwide rather than staying confined to Europe.
My Thoughts
Each week I’m always amazed at the new ways AI is being used. We’re far beyond the point of just personal productivity.
While some of these headlines feel a bit alarming/dystopian, I feel the pros outweigh the cons here.
A thought-to-text headset could be useful for someone who's lost the ability to speak. A gene-edited beagle could let a kid grow up with a dog for the first time. A cheaper Flash model could make an automation project finally get a business off the ground. But every one of these tools asks something of the people building on top of them: know what you're actually trading away for the convenience.
I had fun uncovering all these stories this week, but if you're trying to figure out AI for your organization, you should check out our AIM workshop. It’s a fully funded AI Roadmapping session with our team. We'll walk through what you're running today, what your goals are for the future, what funding is available, and more. If you’re interested in that type of session, reach out to our team here.
Until next time,
Ryan
Now, time for this week’s AI-generated image and the prompt I used to create it.
Create an image of a dystopian world of gene edited people and animals. Create living beings that do not look like today's living beings.
This kind of bummed me out, so I had AI create a less dystopian version and I laughed that it really just added sunshine and more greenery.
