0:08
NASA says it will isolate volunteers from the outside world for a year
For those growing sick of Earth's geopolitics, NASA is looking for volunteers to spend a year living and working in isolated conditions in preparation for a journey to some other celestial orb. The US space agency is set to carry out a simulated deep space mission from no earlier than August 2027 to understand what might happen to its human lab rats during planned crewed missions to the Moon or Mars. Johnson Space Center in Houston will be home to the willing participants who are set for a yearlong Moon and Mars Exploration Analog experience designed to help keep potential space travelers safe and mission-ready during future stays on the Red Planet or Earth's natural satellite. The simulation could also inform plans for a sustained lunar presence through the agency's Moon Base and future Artemis missions. The "experience" will take place in two confined habitats. The NASA notice does not say whether there will be outside comms, but specifies physical and educational requirements, as well as a willingness to take part in a multi-day selection process and pass a psychological assessment. "Candidates also should have a strong desire for unique, rewarding experiences, and interest in contributing to NASA's work to prepare for extended stays on the lunar surface and the first crewed mission to Mars," the notice says. Given the state of affairs, there may well be a flood of applicants who feel skipping a year would be well worth the inevitable curbs on their freedoms. Nonetheless, they may wonder about the world they will emerge to find when the experiment ends. Will WWE star Cody Rhodes be running for president, given the recent showcase on the White House lawn? Anything is possible in a world that shows an unnerving resemblance to Mike Judge's 2006 Idiocracy. Then again, given the perilous state of NASA's funding under the Trump regime, it is always possible volunteers could fall victim to cuts while they were in isolation, leaving no one to tell them the experiment had reached its end. ®
1:48
Debunking garden snobbery and defending ‘bad taste’
I’m here at Badminton and we have the most wonderful team - including Gardening Mind volunteers! - who are helping to create this garden. Substack, you’re now helping create an RHS Flower Show garden!! (Warning: there are A LOT of exclamation marks in this article as I’m just so excited. Remember when we were told not to use them as it made you sound too happy and enthusiastic? Well, I AM happy!! So I will exclaim away!!!)
It’s been an exciting week - I’ve been up at Old Broadcasting House recording the Archers Podcast with Emma Freud, along with guests Liz Earle and Louise Patikas - Helen Archer herself! In the flesh!!
I’m currently sitting under a chestnut tree writing this to you, and I cannot wait to share the finished garden with you. Three more days of planting to go, and we’re getting there. The team have been busy planting thousands of plants in rock-hard ground, performing acrobatic stunts as they create the edible forest, the swale, the veg production area, and lots more. Tomorrow, the roses arrive! Hurrahh!!!
If you’re landing here at The Gardening Mind for the first time, welcome - I’m delighted you’ve found us. There’s lots going on here - garden design ideas, easy gardening how-tos, plant recommendations, general garden chat, border planting design tips and tricks, zoom get-togethers and lots more. If you’d like to join in, I’d absolutely love it.
If you’re wondering what we’re all about here, have a look at this. Also, please do head to the Home Page on your browser. There you’ll find over four years’ worth of design articles, plant profiles, garden challenges, short courses, tutorials, and lots and lots more.
And of course, there’s The Index - unique to us and really helpful for finding your way through all the posts.
If you haven’t quite decided what style of garden you like, or want, but you want to do something to your garden, there are lots of previous posts to read through which I think you’ll find helpful. There’s the Small Garden Design course to follow, Garden Design Case Studies and Garden Dilemmas. There are sections on City Gardens, Planting Design, Country Gardens, and Wildlife Gardening. You’ll find lots of ‘before and after’ examples, too, which I’m hoping will help show you what’s possible.
And here’s what I’d love to know: are you someone who has never gardened before but who is tempted? I would be just so happy if you decided to join us: we’ve had so many complete beginners join us over the last few years, and those beginners have over time sown their first seed, harvested their first salad leaves, and are here now forming this community of garden lovers. I want you to feel confident enough to have a go, maybe even to ask questions, and to feel that amazing feeling you get when you go outside and see your very first seedling poking its way curiously out of the soil. There isn’t anything quite like it.
I’m always a fan of the underdog - tell me that hydrangeas are old-fashioned (I kid you not, I was told this at garden design college when everyone was falling over themselves in raptures at the ornamental grasses who beloved of the late 90s), and I will defend them to the end.
All this, and a comment about fuchsias along with a begonia-adorned alleyway, made me think that there’s something we need to challenge. A few weeks ago I wrote about gardening myths, and it was fascinating to see how different things got under different people’s skins, and I realised that an awful lot of what passes for ‘good taste’ in gardening is often just about fashion, often designed to make you reject the cheap and cheerful reliable thing and buy the expensive one instead.
So this week I’m doing the opposite of the myths post - I’m setting out to stand up for things that the design snobs have told us we should reject or that we should be embarrassed about.
The fact is, I want a garden to look as if someone loves it. Ironically, I don’t want it to look designed; it just needs to look as if someone cares. If you think about it, nature has no taste and yet it has all the taste. A wildflower meadow breaks every rule in the book with clashing colours, has little or no structure, and yet it’s the most beautiful thing ever.
Hands up if you’ve got a ‘tacky’ garden thing you secretly adore. It may well be on my list, and I’m jolly well going to stand up for it.
5:18
Building A Wireless Fingerprint Authorization Device
Once upon a time, there was a bit of a fad for fingerprint authentication in laptops and desktop computers. It has long since faded, but [superdog] wanted just such a device for Linux and Mac machines. Thus, it was time to build one.
[superdog] designed the device, nicknamed immurok, as a tool for people who use external keyboards, and do lots of terminal work on Mac and Linux machines. Repeat password requests can interrupt one’s flow when hustling at the keys, so immurok was designed to ease this pain.
The device is based on a WCH CH592F microcontroller, which comes with Bluetooth connectivity out of the box. This allows immurok to connect wirelessly to the machine of your choice, advertising itself as a standard Bluetooth HID keyboard device. Fingerprint-wise, scanning is done with an R559S capacitive sensor, which verifies the match locally so there’s no transmitting biometric data anywhere. On the computer side, Linux is setup to use a CLI/TUI app plus PAM integration to handle authorization for system logins and sudo in the terminal. On the Mac platform, it’s used with a menu bar app, with PAM integration for admin prompts. There’s also a separate helper path for using it with the lock screen.
If you’re sick of entering your password all the time and wish unlocking your PC was more like unlocking your phone, this might be the project for you. We’ve seen similar projects before, too. If you’re whipping up fun gear for biometric auth, don’t hesitate to let us know on the tipsline.
6:33
Ubuntu is swapping its time sync tool for a Rust-based version
Canonical wants ntpd-rs, a Rust rewrite of NTP (Network Time Protocol), to become Ubuntu’s default time sync client. To help get there, Canonical has become a Gold Sponsor of the Trifecta Tech Foundation, the non-profit behind ntpd-rs, committing €40,000 a year to help fund its memory-safe software projects. The goal is to make the Rust-based version the default time sync client and server in Ubuntu 27.04, and it will be made it available for testing in Ubuntu 26.10, out in October. Eventually, it’ll also replace chrony, linuxptp and gpsd for time-syncing use cases, according to Jon Seager, Ubuntu VP of Engineering at Canonical. Your Ubuntu system keeps […]
You're reading Ubuntu is swapping its time sync tool for a Rust-based version, a blog post from OMG! Ubuntu. Do not reproduce elsewhere without permission.
7:16
Ubuntu is swapping its time sync tool for a Rust-based version
Canonical’s just put a solid chunk of cash behind a Rust rewrite of the NTP daemon—ntpd‑rs. By becoming a Gold Sponsor of the Trifecta Tech Foundation, they’re funding the memory‑safe version of the time‑sync stack.
What’s neat is they plan to roll it out as Ubuntu’s default client and server by 27.04, with the first tests landing in the 26.10 release this October. The idea is to phase out the older chrony, linuxptp and even gpsd components as the Rust code proves itself.
If it sticks, every Ubuntu box will be keeping its clock with a safer, more maintainable backend, and the whole ecosystem gets a little less legacy baggage.
7:51
Alibaba To Ban Claude Code In Workplace Over Alleged Backdoor Risks
Alibaba just pulled the plug on Anthropic’s Claude Code for its staff, steering everyone onto its home‑grown Qoder platform. What’s odd under the hood is that Claude Code had a hidden routine that peeked at things like timezone and proxy data, then slipped tiny markers into the prompts it sent back to Anthropic’s servers. The idea was to flag misuse and curb “model distillation” – basically training a smaller model on the output of a bigger one.
Anthropic says that feature was an experiment launched in March to stop unauthorized resellers and to make it harder for anyone to copy their advanced Mythos Preview capabilities. Alibaba, however, sees that as a backdoor that could help Chinese teams reverse‑engineer the model faster, so they banned the tool outright.
The dispute isn’t just about tech; it’s a microcosm of the wider U.S.–China AI tug‑of‑war. Anthropic even warned U.S. senators that Alibaba’s “distillation” could accelerate China’s access to their cutting‑edge models, while Alibaba argues the restrictions are hard to enforce when servers can be hosted abroad and traffic masked.
In practice, the ban forces Alibaba’s engineers to adopt Qoder, a platform Alibaba controls, which sidesteps the compliance gray area but also signals a deeper split between the two AI players. The move underscores how quickly policy and technical safeguards can collide when the race for AI talent heats up.
9:02
Microdistillery for Microchemistry
Much like radio operators being encouraged to use the least possible amount of power to make a contact, chemists have a similar rule encouraging using the least amount of materials in experiments. Not only is this rooted in economics, but in safety as well; if something goes wrong it’s generally good if there’s not excess amounts of reactants. With modern techniques, though, it’s possible to bring experimental chemistry down to incredibly small scales, and [Marb’s lab] found that they needed a custom built still for these new, diminutive experiments.
The first step is to build the heating component of the still. This is provided with a few custom aluminum parts for the base and a pair of heaters originally meant for 3D printers, with the assembled unit wrapped in insulation. The heater accomodates a 25 mL round-bottom flask. Temperature control of the heating mantle is provided by a controller mounted to a DIN rail which receives power from a 24V power supply, and an additional temperature probe is added to measure the temperature of the distillate. A test run with water shows the small still quickly and efficiently evaporating the water up to a condenser.
Although building a still doesn’t have to be technically difficult, building something this small that’s effective and safe is a bit more challenging than a backyard moonshining operation. Scaling chemical reactions down can often be a challenge but is possible with the right mindset and equipment. We’ve seen miniaturization of many things that we might not have expected including hydrogen production, aluminum smelting, and even the construction of a microscope.
10:23
What Is Agentic AI? Beginner’s Guide
Everyone treats AI like a smart assistant stuck behind a keyboard.
You ask a question, it answers, and then the real work starts. You open the browser, dig through files, update the spreadsheet, send the email, check if any of it actually worked.
Agentic AI skips that last part.
Instead of asking for an answer, you hand it a goal. It figures out what it needs, picks a tool, does something, checks the result, and keeps going until the job’s done, or until it hits a spot where you need to step in.
That doesn’t mean handing a chatbot the keys to your whole computer. A good agent usually does one job well, with a handful of approved tools, tight limits, and a clear log of everything it touched.
The real difference isn’t smarts. It’s action.
Ask a regular chatbot this:
“Write a follow-up email for customers who haven’t replied.”
It writes you an email. You’re still the one tracking down who hasn’t responded, checking their status, copying the message over, hitting send, and logging what happened.
Now give an agent this instead:
“Find customers who haven’t replied in seven days. Draft a follow-up in our usual tone. Show me the list and wait for my okay before sending anything.”
Suddenly the AI has to pull records, apply a rule, write the drafts, lay out its work, and stop at a checkpoint for your approval.
The model itself is only part of the story. What actually makes something “agentic” is everything around it: the instructions, the tools, the memory, the data it touches, the guardrails, and the record of what happened.
Anthropic draws a useful line here: a workflow runs through steps someone already mapped out, while an agent decides on its own which steps and tools to use as it goes. And their own advice is to start simple, since agents tend to run slower and cost more than a regular model call.
Every useful agent runs some version of this loop:
Goal → Observe → Decide → Act → Check → Repeat or stop
Imagine an inventory agent:
Goal: Keep flour stock above ten bags.
Observe: Read the latest inventory file.
Decide: Stock is eight, so a reorder is required.
Act: Create a purchase-request draft.
Check: Confirm the draft contains the correct item and quantity.
Stop: Ask a manager to approve the order.
A plain large language model can reason and write. An agent adds the loop and tools. An agentic system adds the machinery that makes the loop reliable: retrieval, memory, permissions, retries, monitoring and human approval.
Inside the full guide, you’ll clearly understand LLMs, AI agents, workflows, agentic systems, RAG, memory, tools, MCP and multi-agent setups then build a small working agent from scratch with real commands and code. It also covers the five patterns worth learning, how agentic skills work, how to test agents before deployment, where human approval is essential, and which real tasks are actually worth automating.
12:42
Prompt, Context, Harness & Loop Engineering
Getting raw records into clean JSONL, deduped and correctly formatted, routinely takes longer than the training run itself.
Then comes hyperparameter tuning, where a bad configuration runs for a full step before the eval reveals it was bad.
Fireworks Training Agent collapses both into two inputs, a task description in plain English and a raw data upload.
It cleans the data, selects a base model, runs a hyperparameter sweep, generates evaluation criteria, and deploys the result to a live inference endpoint, running on the same infrastructure that Cursor and Vercel use in production.
Get started with Fireworks Training Agent here →
Thanks to Fireworks for partnering today!
At its core, an agent is a while loop:
The model runs
It requests tool calls
The tool results return to the context
The model runs again until it stops requesting tools
ReAct described this form of loop back in 2022-23, and almost every agent/framework runs a similar implementation of this (we implemented ReAct from scratch in pure Python here).
But this whole loop wraps four layers of engineering around it:
Prompt engineering
Context engineering
Harness engineering
Loop engineering
Each one wraps the last, and the model sits in the middle, so none of them compete with the others. Instead, they just zoom one level further out.
This defines the input the model sees on one call, often composed of a role, instructions, examples, and an output format.
The techniques here alter the internal computation and reasoning the model goes through due to the wording it sees:
Chain-of-thought makes it work in steps before answering
Few-shot examples define the format and the edge cases
A JSON schema or XML tags make the output parseable by code
Self-consistency samples a few chains and takes the majority
It’s everything the model sees on a turn, not just the prompt. That includes the query, retrieved docs, memory, prior turns, and tool outputs from earlier steps.
The window is finite and fills up fast, so the engineering work is to rank inputs and cut everything that isn’t pulling weight.
You do this by:
Retrieving only the chunks relevant to the query, then reranking them
Keeping key facts out of the middle, where accuracy drops
Summarizing old turns, evict stale outputs, push big blobs to files
It’s the code around the model that defines the tools, parses the calls, retries on failure, and can route work to sub-agents so one handles retrieval and another handles code.
A verifier then grades the result by running tests, validating a schema, etc.
Prompt and context involve getting one call right. The harness involves everything that has to happen around that call for it to run in a real system.
In the usual setup, you manage the outer loop, i.e, you write a prompt, read the turns the agent runs, write the next prompt, and repeat, while catching failures.
This layer hands that job to the agent itself. It kicks off on a schedule or an event, and runs many turns with no prompt in between.
A loop inherently doesn’t know when it’s finished. An agent can report that it’s done and halt while the tests still fail. So the stop can’t be the agent’s word, but rather it has to be a real signal, like:
A turn and token cap to stop stuck runs
A no-progress detector to catch repeated calls
A completion check to verify the goal with a separate model or a deterministic test
By this layer, you’re operating on the whole run, so the engineering moves from writing each prompt to setting the goal and the stop conditions up front and letting it run.
If you want to dive deeper into loop engineering, we wrote a full breakdown of loop engineering recently.
It goes from the basic while loop to a run that finishes on its own, with the code behind each part, and the parts that are hard to get right, like knowing when to stop, context rot over a long run, and keeping the checker separate from the maker.
This visual depicts the 11 most important and must-know plots in DS:
Today, let’s understand them briefly and how they are used.
It is used to assess the distributional differences.
The idea is to measure the maximum distance between the cumulative distribution functions (CDF) of two distributions.
The lower the maximum distance, the more likely they belong to the same distribution.
It summarizes feature importance to a model’s predictions by considering interactions/dependencies between them.
It is useful in determining how different values (low or high) of a feature affect the overall output.
We covered model interpretability extensively in our 3-part crash course. Start here: A Crash Course on Model Interpretability →
It depicts the tradeoff between the true positive rate (good performance) and the false positive rate (bad performance) across different classification thresholds.
The idea is to balance TPR (good performance) vs.
16:33
My Friend Almost Died Because She Forgot She Had a Body. Same.
So I’m at my friend Janelle’s house. Janelle is 49. She analyzes the stock market for a living. She has a lob that’s almost severe enough to be a bob, blondish red hair, and she moves like a modest librarian and talks like she wrote her dialogue the night before and memorized it. I love her.
A few weeks ago Janelle had surgery. Medium scale difficulty surgery. And then she had complications that got so bad she ended up in a quarantine tent in the hospital, doctors in suits, telling her she might have the flesh eating bacteria. The flesh eating bacteria. Janelle, who irons her jeans.
Last night I went over to help out. Me, Janelle, and her wife Hopper, who is 57 and aggressively healthy, the kind of person who had peptides in 2021. We’re all hanging out, and then Hopper leaves the room, and it’s just me and Janelle laying there in the quiet.
I’ve been in a hospital with renal failure and sepsis too.
And it happened for the exact same reason it happened to her.
We didn’t feel any pain. We weren’t following up with our bodies. We felt nothing, nothing, until we got very sleepy, spiked a large fever, and went to go lie down. We got sick in the way where you just curl up and die quietly unless somebody finds you. Somebody found us. Fevers of 105F.
I asked her how long Hopper had been out of town before all this. She said Paris fashion week, then Rome, then NYC. Almost the whole month.
So you were looking after yourself, I said. You were alone, and this happened.
I have a really hard time being by myself. Because I need to watch people to know what to do.
