Have you heard that author Andy Weir has a new book coming out? Very exciting, we know, and according to a syndicated reading list for Summer 2025, it’s called The Last Algorithm, and it’s a tale of a programmer who discovers a dark and dangerous secret about artificial intelligence. If that seems a little out of sync with his usual space-hacking fare such as The Martian and Project Hail Mary, that’s because the book doesn’t exist, and neither do most of the other books on the list.

The list was published in a 64-page supplement that ran in major US newspapers like the Chicago Sun-Times and the Philadelphia Inquirer. The feature listed fifteen must-read books, only five of which exist, and it’s no surprise that AI is to behind the muck-up. Writer Marco Buscaglia took the blame, saying that he used an LLM to produce the list without checking the results. Nobody else in the editorial chain appears to have reviewed the list either, resulting in the hallucination getting published. Readers are understandably upset about this, but for our part, we’re just bummed that Andy doesn’t have a new book coming out.

In equally exciting but ultimately fake news, we had more than a few stories pop up in our feed about NASA’s recent discovery of urban lights on an exoplanet. AI isn’t to blame for this one, though, at least not directly. Ironically, the rumor started with a TikTok video debunking a claim of city lights on a distant planet. Social media did what social media does, though, sharing only the parts that summarized the false claim and turning a debunking into a bunking. This is why we can’t have nice things.

That wasn’t the only story about distant lights, though, with this report of unexplained signals from two nearby stars. This one is far more believable, coming as it does from retired JPL scientist Richard H. Stanton, who has been using a 30″ telescope to systematically search for optical SETI signals for the past few years. These searches led to seeing two rapid pulses of light from HD 89389, an F-type star located in the constellation Ursa Major. The star rapidly brightened, dimmed, brightened again, then returned to baseline over a fraction of second; the same pattern repeated itself about 4.4 seconds later.

Intrigued, he looked back through his observations and found a similar event from a different star, HD 217014 in Pegasus, four years previously. Interestingly, this G-type star is known to have at least one exoplanet. Stanton made the first observation in 2023, and he’s spent much of the last two years ruling out things like meteor flashes or birds passing through his field of view. More study is needed to figure out what this means, and while it’s clearly not aliens, it’s fun to imagine it could be some kind of technosignature.

And one last space story, this time with the first observation of extra-solar ice. The discovery comes from the James Webb Space Telescope, which caught the telltale signature of ice crystals in a debris ring circling HD 181327, a very young star only 155 light-years away. Water vapor had been detected plenty of times outside our solar system, but not actual ice crystals until now. The ice crystals seem to be coming from collisions between icy bodies in the debris field, an observation that has interesting implications for planetary evolution.

And finally, if like us you’re impressed anytime someone busts out a project with a six-layer PCB design, wait till you get a load of this 124-layer beast. The board comes from OKI Circuit Technologies and is intended for high-bandwidth memory for AI accelerators. The dielectric for each layer is only 125-μm thick, and the board is still only 7.6 mm thick overall. At $4,800 per square meter, it’s not likely we’ll see our friends at JLC PCB offering these anytime soon, but it’s still some pretty cool engineering.

There was a period in the 1990s when it seemed like the personal data assistant (PDA) was going to be the device of the future. If you were lucky you could afford a Psion, a PalmPilot, or even the famous Apple Newton — but to trap the unwary there were a slew of far less capable machines competing for market share.

[Nick Bild] has one of these, branded Rolodex, and in a bid to make using a generative AI less alluring, he’s set it up as the interface to an LLM hosted on a Raspberry Pi 400. This hack is thus mostly a tale of reverse engineering the device’s serial protocol to free it from its Windows application.

Finding the baud rate was simple enough, but the encoding scheme was unexpectedly fiddly. Sadly the device doesn’t come with a terminal because these machines were very much single-purpose, but it does have a memo app that allows transfer of text files. This is the wildly inefficient medium through which the communication with the LLM happens, and it satisfies the requirement of making the process painful.

We see this type of PDA quite regularly in second hand shops, indeed you’ll find nearly identical devices from multiple manufacturers also sporting software such as dictionaries or a thesaurus. Back in the day they always seemed to be advertised in Sunday newspapers and aimed at older people. We’ve never got to the bottom of who the OEM was who manufactured them, or indeed cracked one apart to find the inevitable black epoxy blob processor. If we had to place a bet though, we’d guess there’s an 8051 core in there somewhere.

Want to give an AI the ability to do stuff in Blender? The BlenderMCP addon does exactly that, connecting open-source 3D modeling software Blender to Anthropic’s Claude AI via MCP (Model Context Protocol), which means Claude can directly use Blender and its tools in a meaningful way.

MCP is a framework for allowing AI systems like LLMs (Large Language Models) to exchange information in a way that makes it easier to interface with other systems. We’ve seen LLMs tied experimentally into other software (such as with enabling more natural conversations with NPCs) but without a framework like MCP, such exchanges are bespoke and effectively stateless. MCP becomes very useful for letting LLMs use software tools and perform work that involves an iterative approach, better preserving the history and context of the task at hand.

Using MCP also provides some standardization, which means that while the BlenderMCP project integrates with Claude (or alternately the Cursor AI editor) it could — with the right configuration — be pointed at a suitable locally-hosted LLM instead. It wouldn’t be as capable as the commercial offerings, but it would be entirely private.

Embedded below are three videos that really show what this tool can do. In the first, watch it create a beach scene using assets from a public 3D asset library. In the second, it creates a scene from scratch using a reference image (a ‘low-poly cabin in the woods’), followed by turning that same scene into a 3D environment on a web page, navigable in any web browser.

Back in 2022 we saw Blender connected to an image generator to texture objects, but this is considerably more capable. It’s a fascinating combination, and if you’re thinking of trying it out just make sure you’re aware it relies on allowing arbitrary Python code to be run in Blender, which is powerful but should be deployed with caution.

You’d think a paper from a science team from Carnegie Mellon would be short on fun. But the team behind LegoGPT would prove you wrong. The system allows you to enter prompt text and produce physically stable LEGO models. They’ve done more than just a paper. You can find a GitHub repo and a running demo, too.

The authors note that the automated generation of 3D shapes has been done. However, incorporating real physics constraints and planning the resulting shape in LEGO-sized chunks is the real topic of interest. The actual project is a set of training data that can transform text to shapes. The real work is done using one of the LLaMA models. The training involved converting Lego designs into tokens, just like a chatbot converts words into tokens.

There are a lot of parts involved in the creation of the designs. They convert meshes to LEGO in one step using 1×1, 1×2, 1×4, 1×6, 1×8, 2×2, 2×4, and 2×6 bricks. Then they evaluate the stability of the design. Finally, they render an image and ask GPT-4o to produce captions to go with the image.

The most interesting example is when they feed robot arms the designs and let them make the resulting design. From text to LEGO with no human intervention! Sounds like something from a bad movie.

We wonder if they added the more advanced LEGO sets, if we could ask for our own Turing machine?

Microsoft’s latest Phi4 LLM has 14 billion parameters that require about 11 GB of storage. Can you run it on a Raspberry Pi? Get serious. However, the Phi4-mini-reasoning model is a cut-down version with “only” 3.8 billion parameters that requires 3.2 GB. That’s more realistic and, in a recent video, [Gary Explains] tells you how to add this LLM to your Raspberry Pi arsenal.

The version [Gary] uses has four-bit quantization and, as you might expect, the performance isn’t going to be stellar. If you are versed in all the LLM lingo, the quantization is the way weights are stored, and, in general, the more parameters a model uses, the more things it can figure out.

As a benchmark, [Gary] likes to use what he calls “the Alice question.” In other words, he asks for an answer to this question: “Alice has five brothers and she also has three sisters. How many sisters does Alice’s brother have?” While it probably took you a second to think about it, you almost certainly came up with the correct answer. With this model, a Raspberry Pi can answer it, too.

The first run seems fairly speedy, but it is running on a PC with a GPU. He notes that the same question takes about 10 minutes to pop up on a Raspberry Pi 5 with 4 cores and 8GB of RAM.

We aren’t sure what you’d do with a very slow LLM, but it does work. Let us know what you’d use it for, if anything, in the comments.

There are some other small models if you don’t like Phi4.

“If there’s one thing the Commodore 64 is missing, it’s a large language model,” is a phrase nobody has uttered on this Earth. Yet, you could run one, if you so desired, thanks to [ytm] and the Llama2.c64 project!

[ytm] did the hard work of porting the Llama 2 model to the most popular computer ever made. Of course, as you might expect, the ancient 8-bit machine doesn’t really have the stones to run an LLM on its own. You will need one rather significant upgrade, in the form of 2 MB additional RAM via a C64 REU.

Now, don’t get ahead of things—this is no wide-ranging ChatGPT clone. It’s not going to do your homework, counsel you on your failed marriage, or solve the geopolitical crisis in your local region. Instead, you’re getting the 260 K tinystories model, which is a tad more limited. In [ytm]’s words… “Imagine prompting a 3-year-old child with the beginning of a story — they will continue it to the best of their vocabulary and abilities.”

It might not be supremely capable, but there’s something fun about seeing such a model talking back on an old-school C64 display. If you’ve been hacking away at your own C64 projects, don’t hesitate to let us know. We certainly can’t get enough of them!

Thanks to [ytm] for the tip!

Looks like the Simpsons had it right again, now that an Australian radio station has been caught using an AI-generated DJ for their midday slot. Station CADA, a Sydney-based broadcaster that’s part of the Australian Radio Network, revealed that “Workdays with Thy” isn’t actually hosted by a person; rather, “Thy” is a generative AI text-to-speech system that has been on the air since November. An actual employee of the ARN finance department was used for Thy’s voice model and her headshot, which adds a bit to the creepy factor.

The discovery that they’ve been listening to a bot for months apparently has Thy’s fans in an uproar, although we suspect that the media doing the reporting is probably more exercised about this than the general public. Radio stations have used robo-jocks for the midday slot for ages, albeit using actual human DJs to record patter to play between tunes and commercials. Anyone paying attention over the last few years probably shouldn’t be surprised by this development, and we suspect similar disclosures will be forthcoming across the industry now that the cat’s out of the bag.

Also from the world of robotics, albeit the hardware kind, is this excellent essay from Brian Potter over at Construction Physics about the sad state of manual dexterity in humanoid robots. The whole article is worth reading, not least for the link to a rogue’s gallery of the current crop of humanoid robots, but briefly, the essay contends that while humanoid robots do a pretty good job of navigating in the world, their ability to do even the simplest tasks is somewhat wanting.

Brian’s example of unwrapping and applying a Band-Aid, a task that any toddler can handle, as being unimaginably difficult for any current robot to handle is quite apt. He attributes the gap in abilities between gross movements and fine motor control partly to hardware and partly to software. We think the blame skews more to the hardware side; while the legs and torso of the typical humanoid robot offer a lot of real estate for powerful actuators, squeezing that much equipment into a hand approximately the size of a human’s is a tall order. These problems will likely be overcome, of course, and when they do, Brian’s helpful list of “Dexterity Evals” or something similar will act as a sort of Turing test for robot dexterity. Although the day a humanoid robot can start a new roll of toilet paper without tearing the first sheet is the day we head for the woods.

We recently did a story on the use of nitrogen-vacancy diamonds as magnetic sensors, which we found really exciting because it’s about the simplest way we’ve seen to play with quantum physics at home. After that story ran, eagle-eyed reader Kealan noticed that Brian over at the “Real Engineering” channel on YouTube had recently run a video on anti-submarine warfare, which includes the uses of similar quantum magnetometers to detect submarines. The magnetometers in the video are based on the Zeeman effect and use laser-pumped helium atoms to detect tiny variations in the Earth’s magnetic field due to large ferrous objects like submarines. Pretty cool video; check it out.

And finally, if you have the slightest interest in civil engineering you’ve got to check out Animagraff’s recent 3D tour of the insides of Hoover Dam. If you thought a dam was just a big, boring block of concrete dumped in the middle of a river, think again. The video is incredibly detailed and starts with accurate 3D models of Black Canyon before the dam was built. Every single detail of the dam is shown, with the “X-ray views” of the dam with the surrounding rock taken away being our favorite bit — reminds us a bit of the book Underground by David Macaulay. But at the end of the day, it’s the enormity of Hoover Dam that really comes across in this video. The way that the structure dwarfs the human-for-scale included in almost every sequence is hard to express — megalophobics, beware. We were also floored by just how much machinery is buried in all that concrete. Sure, we knew about the generators, but the gates on the intake towers and the way the spillways work were news to us. Highly recommended.

While tools like CRISPR have blown the field of genome hacking wide open, being able to predict what will happen when you tinker with the code underlying the living things on our planet is still tricky. Researchers at Stanford hope their new Evo 2 DNA generative AI tool can help.

Trained on a dataset of over 100,000 organisms from bacteria to humans, the system can quickly determine what mutations contribute to certain diseases and what mutations are mostly harmless. An “area we are hopeful about is using Evo 2 for designing new genetic sequences with specific functions of interest.”

To that end, the system can also generate gene sequences from a starting prompt like any other LLM as well as cross-reference the results to see if the sequence already occurs in nature to aid in predicting what the sequence might do in real life. These synthetic sequences can then be made using CRISPR or similar techniques in the lab for testing. While the prospect of building our own Moya is exciting, we do wonder what possible negative consequences could come from this technology, despite the hand-wavy mention of not training the model on viruses to “to prevent Evo 2 from being used to create new or more dangerous diseases.”

We’ve got you covered if you need to get your own biohacking space setup for DNA gels or if you want to find out more about powering living computers using electricity. If you’re more curious about other interesting uses for machine learning, how about a dolphin translator or discovering better battery materials?

This week, the hackerverse was full of “vibe coding”. If you’re not caught up on your AI buzzwords, this is the catchy name coined by [Andrej Karpathy] that refers to basically just YOLOing it with AI coding assistants. It’s the AI-fueled version of typing in what you want to StackOverflow and picking the top answers. Only, with the current state of LLMs, it’ll probably work after a while of iterating back and forth with the machine.

It’s a tempting vision, and it probably works for a lot of simple applications, in popular languages, or generally where the ground is already well trodden. And where the stakes are low, as [Al Williams] pointed out while we were talking about vibing on the podcast. Can you imagine vibe-coded ATM software that probably gives you the right amount of money? Vibe-coding automotive ECU software?

While vibe coding seems very liberating and hands-off, it really just changes the burden of doing the coding yourself into making sure that the LLM is giving you what you want, and when it doesn’t, refining your prompts until it does. It’s more like editing and auditing code than authoring it. And while we have no doubt that a stellar programmer like [Karpathy] can verify that he’s getting what he wants, write the correct unit tests, and so on, we’re not sure it’s the panacea that is being proclaimed for folks who don’t already know how to code.

Vibe coding should probably be reserved for people who already are expert coders, and for trivial projects. Just the way you wouldn’t let grade-school kids use calculators until they’ve mastered the basics of math by themselves, you shouldn’t let junior programmers vibe code: It simultaneously demands too much knowledge to corral the LLM, while side-stepping any of the learning that would come from doing it yourself.

And then there’s the security side of vibe coding, which opens up a whole attack surface. If the LLM isn’t up to industry standards on simple things like input sanitization, your vibed code probably shouldn’t be anywhere near the Internet.

So should you be vibing? Sure! If you feel competent overseeing what [Dan] described as “the worst summer intern ever”, and the states are low, then it’s absolutely a fun way to kick the tires and see what the tools are capable of. Just go into it all with reasonable expectations.

Will a 486 run Crysis? No, of course not. Will it run a large language model (LLM)? Given the huge buildout of compute power to do just that, many people would scoff at the very notion. But [Yeo Kheng Meng] is not many people.

He has set up various DOS computers to run a stripped down version of the Llama 2 LLM, originally from Meta. More specifically, [Yeo Kheng Meng] is implementing [Andreq Karpathy]’s Llama2.c library, which we have seen here before, running on Windows 98.

Llama2.c is a wonderful bit of programming that lets one inference a trained Llama2 model in only seven hundred lines of C. It it is seven hundred lines of modern C, however, so porting to DOS 6.22 and the outdated i386 architecture took some doing. [Yeo Kheng Meng] documents that work, and benchmarks a few retrocomputers. As painful as it may be to say — yes, a 486 or a Pentium 1 can now be counted as “retro”.

The models are not large, of course, with TinyStories-trained  260 kB model churning out a blistering 2.08 tokens per second on a generic 486 box. Newer machines can run larger models faster, of course. Ironically a Pentium M Thinkpad T24 (was that really 21 years ago?) is able to run a larger 110 Mb model faster than [Yeo Kheng Meng]’s modern Ryzen 5 desktop. Not because the Pentium M is going blazing fast, mind you, but because a memory allocation error prevented that model from running on the modern CPU. Slow and steady finishes the race, it seems.

This port will run on any 32-bit i386 hardware, which leaves the 16-bit regime as the next challenge. If one of you can get an Llama 2 hosted locally on an 286 or a 68000-based machine, then we may have to stop asking “Does it run DOOM?” and start asking “Will it run an LLM?”

[Geoffrey Litt] shows that getting an effective digital assistant that’s tailored to one’s own needs just needs a little DIY, and thanks to the kinds of tools that are available today, it doesn’t even have to be particularly complex. Meet Stevens, the AI assistant who provides the family with useful daily briefs. The back end? Little more than one SQLite table and a few cron jobs.

Every day, Stevens sends a daily brief via Telegram that includes calendar events, appointments, weather notes, reminders, and even a fun fact for the day. Stevens isn’t just send-only, either. Users can add new entries or ask questions about items through Telegram.

It’s rudimentary, but [Geoffrey] already finds it far more useful than Siri. This is unsurprising, as it has been astutely observed that big tech’s digital assistants are designed to serve their makers rather than their users. Besides, it’s also fun to have the freedom to give an assistant its own personality, something existing offerings sorely lack.

Architecture-wise, the assistant has a notebook (the single SQLite table) that gets populated with entries. These entries come from things like reading family members’ Google calendars, pulling data from a public weather API, processing delivery notices from the post office, and Telegram conversations. With a notebook of such entries (along with a date the entry is expected to be relevant), generating a daily brief is simple. After all, LLMs (Large Language Models) are amazingly good at handling and formatting natural language. That’s something even a locally-installed LLM can do with ease.

[Geoffrey] says that even this simple architecture is super useful, and it’s not even a particularly complex system. He encourages anyone who’s interested to check out his project, and see for themselves how useful even a minimally-informed assistant can be when it’s designed with ones’ own needs in mind.

Giskard is an open-source AI model quality testing tool that helps data scientists and engineers build safer, more reliable AI systems. The platform was built by AI engineers for AI engineers. It’s completely open source and designed to help teams and developers build more robust, trustworthy AI models. To use the platform, you can get […]

Source

Faraday.dev lets you easily run open-source LLMs (chatbots) on your computer. Once you’ve got the program and AI models installed, no internet connection is required to use and interact with the AI LLMs. Faraday.dev supports a wide range of LLaMA-based models, including WizardLM, GPT4-x-Alpaca, Vicuna, Koala, Open Assistant, PygmalionAI, and more. You have the option […]

Source

Oobabooga is an open-source Gradio web UI for large language models that provides three user-friendly modes for chatting with LLMs: a default two-column view, a notebook-style interface, and a chat interface. This flexibility allows you to interact with the AI models in a way that best suits your needs, whether it’s for writing, analysis, question-answering, […]

Source

Code Llama is a suite of large language models released by Meta AI for generating and enhancing code. It includes foundation models for general coding, Python specializations, and models tailored for following instructions. Key features include state-of-the-art performance, code infilling, large context support up to 100K tokens, and zero-shot ability to follow instructions for programming […]

Source

ChatGLM-6B is an open-source, bilingual conversational AI LLM based on the General Language Model (GLM) framework. It has 6.2 billion parameters and can be deployed locally with only 6GB of GPU memory. This model allows for natural language processing in both Chinese and English, question answering, task-oriented dialogue, and easy integration via API and demo […]

Source

Perplexity AI is an AI chat and search engine that uses advanced technology to provide direct answers to your queries. It delivers accurate answers using large language models and even includes links to citations and related topics. It is available for free via web browser and also on mobile via the Apple App Store. Using […]

Source

Codestral is a powerful 22B parameter AI model from Mistral AI. This open-weight model is designed specifically for code generation across over 80 programming languages including Python, Java, C++, JavaScript and more. Codestral offers impressive performance, outperforming other models on benchmarks like HumanEval and RepoBench with its large 32k token context window. The model is […]

Source

Langtail is a platform that helps you develop and deploy LLM-powered applications faster. It provides tools for prompt engineering, testing, observability, and deployment – all in one place. You can collaborate with your team, iterate quickly, and get your LLM apps to production with confidence.

Source