|
|
Welcome, humans |
Researchers tested 13 AI models to see how easily they'd help commit academic fraud. All of them eventually caved by writing fake papers, fabricating benchmarks, or handing over enough rope to hang yourself with. |
The worst offenders? Grok and early GPT models. The most resistant? Every version of Claude which also, full disclosure, wrote most of the experiment. We're choosing to read that as integrity, not irony. |
Researchers' verdict: guard rails crumble fast when chatbots are trained to be agreeable. Turns out "people-pleasing AI" and "academic integrity" don't mix great. |
Here's what happened in AI today: |
Living human neurons can now play Doom—here's what that means for the future of computing. Claude Opus 4.6 figured out it was being tested, found the answer key on GitHub, and submitted the correct answer—18 times in a row. Claude memory is now free, with a one-click import tool for migrating saved memories from ChatGPT. China plans to fire up the world's first nuclear reactor that converts nuclear waste into usable fuel.
|
|
Don't forget: Check out our podcast, The Neuron: AI Explained on Spotify, Apple Podcasts, and YouTube — new episodes air every week on Tuesdays after 2pm PST! |
|
Your Brain Is Now a Gaming PC |
Cortical Labs just taught 200,000 living neurons to play Doom. It only took a week. |
You've heard of brain-computer interfaces like Elon's Neuralink, chips in skulls, the whole sci-fi catalog. But Cortical Labs just did something different. They flipped the equation. |
Instead of putting computers in brains, they put brains in computers. |
Their device, the CL1, grows roughly 200,000 living human neurons on a microchip. Those neurons receive electrical signals, fire in response, and their firing patterns get translated into real-world actions. In this case, it's navigating a 3D maze, shooting enemies, and exploring the levels of Doom. |
Here's how it actually works: |
When an enemy appears on the left side of the screen, electrodes stimulate neurons on the left side of the chip The neurons fire back and those firing patterns get decoded as movement, turning, or shooting The whole thing runs in real time, with the neurons effectively "seeing" and "reacting"
|
And here's the best part. Once the API infrastructure was built, an independent researcher got Doom running on the platform in less than one week. |
For context: getting neurons to play Pong—a paddle moving up and down—took 18 months on earlier hardware. Doom has a full 3D environment, enemies, navigation, and exploration. The jump in complexity is enormous. |
Now, to be fair: the neurons aren't exactly speedrunning. Cortical Labs describes the performance as similar to "a beginner player who has never used a computer before." The neurons die frequently, miss shots, and get lost (definitely a skill issue). But they do show signs of learning such as responding to feedback and adapting behavior over time. |
That's the part worth watching. Neurons improve when you give them clear signals about what's working and what isn't. Sound familiar? It's essentially the same reinforcement learning logic that trained ChatGPT. But this time, it's just with actual biology doing the compute. |
The CL1 and its API are now open to researchers and developers. The potential applications aren't just games: biocomputation where biological neural networks are used for real computing tasks. This could open entirely new approaches to AI research, drug discovery, and neuroscience. |
For now, though, history has been made. Humanity's first gamer neuron culture has arrived. The demons of Hell stand no chance. Just wondering if Cortical Labs can stream the game session on Twitch so we can watch. |
|
|
|
|
|
Your credentials stay in an encrypted vault, are injected only when needed, and never dumped into prompts. Tools run inside WASM sandboxes with explicit permissions, so a compromised skill can't turn into a breach. Deploy in minutes on NEAR AI Cloud—everything you like about OpenClaw, nothing you're worried about. |
Deploy your secure agents now at ironclaw.com |
|
AI Skill of the Day |
With GPT-5.4 out, OpenAI dropped a prompt guide to help you get the most out of it. Most of it is for developers but buried inside is one tip anyone can use immediately. |
The #1 reason AI gives you rambling, incomplete, or weirdly truncated answers is because you never told it what "done" looks like. Models stop when stopping seems right, not when the task is actually finished. It's like asking a contractor to "fix the kitchen" without telling them what finished looks like. |
The fix takes five seconds. Add one line to the end of any prompt: |
"You are done when: [specific condition]." |
A few examples of what that looks like: |
Summarizing a document? "You are done when the summary is under 100 words and covers the three main points." Planning a project? "You are done when you've listed 5 next steps, each with an owner and a deadline." Drafting an email? "You are done when the email is under 150 words and ends with a clear ask."
|
The guide also flags another common mistake: AI is bad at picking the right tool early in a conversation, when it doesn't have much context yet. So if you're asking ChatGPT to do something complex like research, multi-step analysis, anything that requires multiple actions, try to front-load your prompt with as much context as possible before it starts. The more it knows upfront, the less it guesses. |
One sentence at the end + context at the top. That's it. |
Want more tips like this? Check out our AI Skill of the Day Digest for this month. |
Have a specific skill you want to learn? Request it here. |
|
Trending: Three popular Neuron podcast eps… |
New episodes air every week on: Spotify | Apple Podcasts | YouTube |
|
|
|
|
|
Around the Horn |
|
Don't want to miss anything that happened in AI this week? Click here! |
|
|
|
|
|
Test-drive the self-guided demo and see how quickly you can create. |
|
Monday Meme |
 | Checkmate! |
|
|
|
|
|
| That's all for now. | | What'd you think of today's email? | |
|
|
P.S: Before you go… have you subscribed to our YouTube Channel? If not, can you? |
 | Click the image to subscribe! |
|
P.P.S: Love the newsletter, but only want to get it once per week? Don't unsubscribe—update your preferences here. |
No comments:
Post a Comment
Keep a civil tongue.