I ricercatori hanno svelato un nuovo modulo di linguaggio AI chiamato xTrimoPGLM in grado sia di comprendere che di creare sequenze proteiche utilizzando un unico approccio di apprendimento. Questa soluzione offre un modo fondamentalmente nuovo di lavorare con i dati biologici, a livello di sequenze di amminoacidi, come se si trattasse di testo in linguaggio naturale.

I modelli precedentemente esistenti che studiavano le proteine ​​si basavano su diversi metodi di pre-addestramento: alcuni ripristinavano le regioni mancanti (autocodifica), altri prevedevano l’amminoacido successivo (autoregressivo).

Tuttavia, ognuno di loro era bravo solo in un compito: comprendere la struttura di una proteina o generarne di nuove. xTrimoPGLM combina entrambe le strategie ed è il primo a imparare da un obiettivo comune, abbracciando entrambi gli approcci contemporaneamente. Ciò ha consentito al modello di raggiungere dimensioni senza precedenti: 100 miliardi di parametri e un trilione di token di addestramento.

I risultati sono impressionanti: xTrimoPGLM ha ottenuto i risultati migliori in 18 diverse attività di analisi proteica, tra cui classificazione, proprietà e previsione delle interazioni. Inoltre, il modello consente di comprendere la struttura delle proteine ​​a livello atomico e viene utilizzato come base per un nuovo modello 3D che prevede la struttura delle proteine ​​con una precisione superiore a strumenti simili basati su modelli linguistici.

Ma non solo analisi: il modello può anche creare. È in grado di generare nuove proteine ​​seguendo i principi delle sequenze naturali. Dopo un ulteriore addestramento su set di dati specializzati, è addirittura possibile eseguire una generazione mirata con proprietà specifiche, il che apre la strada allo sviluppo di nuovi farmaci ed enzimi.

I pesi del modello e i set di dati sono già pubblicati sulla piattaforma HuggingFace e sono a disposizione dei ricercatori, rendendo xTrimoPGLM un contributo significativo allo sviluppo di modelli di base proteici e ampliando gli orizzonti dell’intelligenza artificiale nelle scienze della vita.

L'articolo Capacità generative: xTrimoPGLM crea nuove sequenze proteiche proviene da il blog della sicurezza informatica.

Water is an excellent coolant, but the flip side is that it is also an excellent solvent. This, in short, is why any water cooling loop is also a prime candidate for an interesting introduction to the galvanic metal series, resulting in severe corrosion that commences immediately. In a recent video by [der8aer], this issue is demonstrated using a GPU cold plate. The part is made out of nickel-plated copper and features many small channels to increase surface area with the coolant.
The surface analysis of the sample cold plate after a brief exposure to distilled water shows the deposited copper atoms. (Credit: der8auer, YouTube)
Theoretically, if one were to use distilled water in a coolant loop that contains a single type of metal (like copper), there would be no issue. As [der8auer] points out, fittings, radiators, and the cooling block are nearly always made of various metals and alloys like brass, for example. This thus creates the setup for galvanic corrosion, whereby one metal acts as the anode and the other as a cathode. While this is desirable in batteries, for a cooling loop, this means that the water strips metal ions off the anode and deposits them on the cathode metal.

The nickel-plated cold plate should be immune to this if the plating were perfect. However, as demonstrated in the video, even a brief exposure to distilled water at 60°C induced strong galvanic corrosion. Analysis in an SEM showed that the imperfect nickel plating allowed copper ions to be dissolved into the water before being deposited on top of the nickel (cathode). In a comparison with another sample that had a coolant with corrosion inhibitor (DP Ultra) used, no such corrosion was observed, even after much longer exposure.

This DP Ultra coolant is mostly distilled water but has glycol added. The glycol improves the pH and coats surfaces to prevent galvanic corrosion. The other element is benzotriazole, which provides similar benefits. Of course, each corrosion inhibitor targets a specific environment, and there is also the issue with organic films forming, which may require biocides to be added. As usual, water cooling has more subtlety than you’d expect.

youtube.com/embed/7pIpKetQlZs?…

hackaday.com/2025/04/20/preven…

China played host to what, presumably, was the world’s first robot and human half-marathon. You can check out the action and the Tiangong Ultra robot that won in the video below. The event took place in Beijing and spanned 21.1 km. There was, however, a barrier between lanes for humans and machines.

The human rules were the same as you’d expect, but the robots did need a few concessions, such as battery swap stops. The winning ‘bot crossed the finish line in just over 160 minutes. However, there were awards for endurance, gait design, and design innovation.

Humans still took the top spots, though. We also noted that some of the robots had issues where they lost control or had other problems. Even the winner fell down once and had three battery changes over the course.

Of the 21 robots that started, only six made the finish line. We don’t know how many of the 12,000 humans finished, but we are pretty sure it was more than six, so we don’t think runners have to worry about robot overlords yet. But they’re getting better all the time.

youtube.com/embed/BYmCxYTbg50?…

hackaday.com/2025/04/20/china-…

We appear to be edging ever closer to a solid statement of “We are not alone” in the universe with this week’s announcement of the detection of biosignatures in the atmosphere of exoplanet K2-18b. The planet, which is 124 light-years away, has been the focus of much attention since it was discovered in 2015 using the Kepler space telescope because it lies in the habitable zone around its red-dwarf star. Initial observations with Hubble indicated the presence of water vapor, and follow-up investigations using the James Webb Space Telescope detected all sorts of goodies in the atmosphere, including carbon dioxide and methane. But more recently, JWST saw signs of dimethyl sulfide (DMS) and dimethyl disulfide (DMDS), organic molecules which, on Earth, are strongly associated with biological processes in marine bacteria and phytoplankton.

The team analyzing the JWST data says that the data is currently pretty good, with a statistical significance of 99.7%. That’s a three-sigma result, and while it’s promising, it’s not quite good enough to seal the deal that life evolved more than once in the universe. If further JWST observations manage to firm that up to five sigma, it’ll be the most important scientific result of all time. To our way of thinking, it would be much more significant than finding evidence of ancient or even current life in our solar system, since cross-contamination is so easy in the relatively cozy confines of the Sun’s gravity well. K2-18b is far enough away from our system as to make that virtually impossible, and that would say a lot about the universality of biochemical evolution. It could also provide an answer to the Fermi Paradox, since it could indicate that the galaxy is actually teeming with life but under conditions that make it difficult to evolve into species capable of making detectable techno-signatures. It’s hard to build a radio or a rocket when you live on a high-g water world, after all.

Closer to home, there’s speculation that the famous Antikythera mechanism may not have worked at all in its heyday. According to researchers from Universidad Nacional de Mar del Plata in Argentina, “the world’s first analog computer” could not have worked due to the accumulated mechanical error of its gears. They blame this on the shape of the gear teeth, which appear triangular on CT scans of the mechanism, and which they seem to attribute to manufacturing defects. Given the 20-odd centuries the brass-and-iron device spent at the bottom of the Aegean Sea and the potential for artifacts in CT scans, we’re not sure it’s safe to pin the suboptimal shape of the gear teeth on the maker of the mechanism. They also seem to call into question the ability of 1st-century BCE craftsmen to construct a mechanism with sufficient precision to serve as a useful astronomical calculator, a position that Chris from Clickspring has been putting the lie to with his ongoing effort to reproduce the Antikythera mechanism using ancient tools and materials. We’re keen to hear what he has to say about this issue.

Speaking of questionable scientific papers, have you heard about “vegetative electron microscopy”? It’s all the rage, having been mentioned in at least 22 scientific papers recently, even though no such technique exists. Or rather, it didn’t exist until around 2017, when it popped up in a couple of Iranian scientific papers. How it came into being is a bit of a mystery, but it may have started with faulty scans of a paper from the 1950s, which had the terms “vegetative” and “electron microscopy” printed in different columns but directly across from each other. That somehow led to the terms getting glued together, possibly in one of those Iranian papers because the Farsi spelling of “vegetative” is very similar to “scanning,” a much more sensible prefix to “electron microscopy.” Once the nonsense term was created, it propagated into subsequent papers of dubious scientific provenance by authors who didn’t bother to check their references, or perhaps never existed in the first place. The wonders of our AI world never cease to amaze.

And finally, from the heart of Silicon Valley comes a tale of cyber hijinks as several crosswalks were hacked to taunt everyone’s favorite billionaires. Twelve Palo Alto crosswalks were targeted by persons unknown, who somehow managed to gain access to the voice announcement system in the crosswalks and replaced the normally helpful voice messages with deep-fake audio of Elon Musk and Mark Zuckerberg saying ridiculous but plausible things. Redwood City and Menlo Park crosswalks may have also been attacked, and soulless city officials responded by disabling the voice feature. We get why they had to do it, but as cyberattacks go, this one seems pretty harmless.

youtube.com/embed/Uy1oNvsUQ0o?…

hackaday.com/2025/04/20/hackad…

Forth is popular on small computers because it is simple to implement, yet quite powerful. But what happens when you really need to shrink it? Well, if your target is the 6502, there’s milliForth-6502.

This is a port of milliForth, which is a fork of sectorforth. The sectorforth project set the standard, implementing a Forth so small it could fit in a 512-byte boot sector. The milliForth project took sectorforth and made it even smaller, weighing in at only 336 bytes. However, both milliForth and sectorforth are for the x86 architecture. With milliForth-6502, [Alvaro G. S. Barcellos] wanted to see how small he could make a 6502 implementation.

So how big is the milliForth-6502 binary? Our tests indicate: 1,110 bytes. It won’t quite fit in a boot sector, but it’s pretty small!

Most of the code for milliForth-6502 is assembly code in sector-6502.s. This code is compiled using tools from the cc65 project. To run the code lib6502 is used for 6502 emulation.

Emulation is all well and good as far as it goes, especially for development and testing, but we’d love to see this code running on a real 6502. Even better would be a 6502 built from scratch! If you get this code running we’d love to hear how it went!

youtube.com/embed/6P4tEYLEhU8?…

hackaday.com/2025/04/20/millif…

Le e-mail interne inviate dai dipendenti di Snapchat, la popolare app di messaggistica fotografica e video, dimostrano che l’azienda ha deliberatamente danneggiato adolescenti e giovani adulti, ovvero il pubblico a cui si rivolge.

Lo psicologo sociale e autore del best-seller Generation Anxious Jonathan Haidt ha pubblicato la scorsa settimana un articolo su Substack in cui detto che Snap Inc. riceve ogni mese migliaia di segnalazioni di “sextortion”, ovvero estorsione sessuale.

Insieme al ricercatore Zach Rush, ha presentato decine di estratti da rapporti interni, studi, e-mail e dichiarazioni pubbliche in cui i dipendenti e i consulenti di Snap riconoscono che Snapchat danneggia i minori. Secondo gli autori, queste confessioni possono essere raggruppate in cinque aree: uso compulsivo, accesso a droghe e armi, distribuzione di materiale pedopornografico e casi di rapporti intimi forzati, cyberbullismo e consapevolezza di danni e violazioni dell’età senza azioni appropriate da parte dell’azienda.

Così come nel caso Per quanto riguarda TikTok, Haidt e Rush affermano che i dipendenti di Snap erano ben consapevoli della portata del problema, ma non stavano adottando le misure necessarie. Alcune di queste problematiche sono note da tempo, come la funzione “streak“, che conta il numero di giorni consecutivi in ​​cui gli utenti condividono foto.

Ciò crea un falso senso di impegno e attaccamento negli adolescenti, e la perdita di questo “lato positivo” può essere percepita quasi come la perdita di un’amicizia. Allo stesso tempo, l’applicazione in sé rimane un mezzo di comunicazione superficiale: non garantisce un contatto reale e non sostituisce l’interazione dal vivo.

È noto che molti ex dirigenti di colossi della tecnologia, tra cui quelli che hanno partecipato alle riprese del documentario “The Social Dilemma“, non permettono ai propri figli di utilizzare gli smartphone. Sanno fin troppo bene cosa devono affrontare gli utenti, soprattutto quelli più giovani.

E se gli inventori di questi sistemi si rifiutano di affidarli ai propri figli, perché dovrebbero farlo i genitori comuni?

Gli autori del rapporto ritengono che sia giunto il momento di avviare una riflessione seria sulla dipendenza digitale e sulla responsabilità delle aziende tecnologiche.

Forse, con l’aumento delle cause legali e il cambiamento dell’opinione pubblica, più persone inizieranno a considerare il prezzo che pagano per essere costantemente presenti nel mondo digitale.

L'articolo Snapchat sapeva tutto: come l’app ha danneggiato milioni di adolescenti proviene da il blog della sicurezza informatica.

Despite the best efforts of the RepRap community over the last twenty years, self-replicating 3D printers have remained a stubbornly elusive goal, largely due to the difficulty of printing electronics. [Brian Minnick]’s fully-printed 3D printer could eventually change that, and he’s already solved an impressive number of technical challenges in the process.

[Brian]’s first step was to make a 3D-printable motor. Instead of the more conventional stepper motors, he designed a fully 3D-printed 3-pole brushed motor. The motor coils are made from solder paste, which the printer applies using a custom syringe-based extruder. The paste is then sintered at a moderate temperature, resulting in traces with a resistivity as low as 0.001 Ω mm, low enough to make effective magnetic coils.

Brushed motors are less accurate than stepper motors, but they do have a particularly useful advantage here: their speed can be controlled simply by varying the voltage. This enables a purely electromechanical control system – no microcontroller on this printer! A 3D-printed data strip encodes instructions for the printer as holes in a plastic sheet, which open and close simple switches in the motor controller. These switches control the speed, direction, and duration of the motors’ movement, letting the data strip encode motion vectors.

Remarkably, the hotend on this printer is also 3D-printed. [Brian] took advantage of the fact that PEEK’s melting point increases by about 110 ℃ when it’s annealed, which should allow an annealed hotend to print itself. So far it’s only extruded PLA, but the idea seems sound.

The video below the break shows a single-axis proof of concept in action. We haven’t been able to find any documentation of a fully-functional 3D printer, but nevertheless, it’s an impressive demonstration. We’ve covered similar printers before, and if you make progress in this area, be sure to send us a tip.

youtube.com/embed/xiygZo0YxBw?…

hackaday.com/2025/04/20/the-mo…

Entry-level oscilloscopes are a great way to get some low-cost instrumentation on a test bench, whether it’s for a garage lab or a schoolroom. But the cheapest ones are often cheap for a reason, and even though they work well for the price they won’t stand up to more advanced equipment. But missing features don’t have to stay missing forever, as it’s possible to augment them to get some of these features. [Tommy’s] project shows you one way to make a silk purse from a sow’s ear, at least as it relates to oscilloscopes.

Most of the problem with these lower-cost tools is their low precision due to fewer bits of analog-digital conversion. They also tend to be quite noisy, further lowering the quality of the oscilloscope. [Tommy] is focusing his efforts on the DSO138-mini, an oscilloscope with a bandwidth of 100 kHz and an effective resolution of 10 bits. The first step is to add an anti-aliasing filter to the input, which is essentially a low-pass filter that removes high frequency components of the signal, which could cause a problem due to the lower resolution of the device. After that, digital post-processing is done on the output, which removes noise caused by the system’s power supply, among other things, and essentially acts as a second low-pass filter.

In part 2 of the project, [Tommy] demonstrates the effectiveness of these two methods with experimental data, showing that a good percentage of the noise on a test signal has been removed from the output. All the more impressive here is that the only additional cost besides the inexpensive oscilloscope itself is for a ceramic capacitor that costs around a dollar. We were also impressed: [Tommy] is a junior in high school!

Presumably, you could apply these techniques to other inexpensive equipment, like this even cheaper oscilloscope based on the ESP32.

hackaday.com/2025/04/20/low-co…

In this series of 23 YouTube videos [Rich] puts the AMD Zynq-7000 SoC through its paces by building a development board from the ground up to host it along with its peripherals. The Zynq is part FPGA and part CPU, and while it has been around for a while, we don’t see nearly as many projects about it as we’d like.

Rich covers everything from the power system to HDMI, USB, DDR RAM, and everything in between. By the end, he’s able to boot PetaLinux.

The Zynq SoC includes an ARM Cortex-A9 Based APU and an Artix-7 FPGA (or a Kintex-7 FPGA on higher models). In case you missed it, Xilinx was recently acquired by AMD, which is why you might have remembered this as a Xilinx part.

We’ve heard from [Rich] before. Back in 2021 we saw his Arduino Brings USB Mouse To Homebrew Computer. Don’t miss his follow-up playlist: Building on my Zynq-7000 in which he takes his Zynq-7000 board even further.

If you’re interested in FPGA technology but need something more easy going to get you started, be sure to check out how to build a 6809 CPU on an FPGA. Or, if you need something even simpler, report for boot camp.

youtube.com/embed/jbG5KUbGaGE?…

Thanks to [Alex] for the tip!

hackaday.com/2025/04/20/buildi…

WASHINGTON (AP) — The State Department has shut down its office that sought to deal with misinformation and disinformation that Russia, China and Iran have been accused of spreading.

Secretary of State Marco Rubio said in a statement Wednesday that he had closed what had been known as the Global Engagement Center because it had taken actions to restrict freedom of speech in the United States and elsewhere.

The center has been a frequent target of criticism from conservatives for calling out media and online reports that it said are biased or untruthful. At times, it has identified U.S. websites and social media accounts that it argued were amplifying misinformation, particularly related to the Russia-Ukraine war.

“It is the responsibility of every government official to continuously work to preserve and protect the freedom for Americans to exercise their free speech,” Rubio said, charging that the office worked “to actively silence and censor the voices of Americans they were supposed to be serving.”

Rubio said that is “antithetical” to the principles that “we should be upholding.”

Congress had supported the center’s work, as had previous State Department leadership.

“This is a deeply misleading (and) unserious portrayal of an organization focused on identifying foreign — primarily Russian — disinformation ops,” former State Department spokesman Ned Price said in a post on X.

dicorinto.it/articoli/the-stat…

When we say non-planar slicing is for the birds, we mean [Joshua Bird], who demonstrates the versatility of his new non-planar S4-Slicer by printing a Benchy upside down with the “Core R-Theta” printer we have featured here before.
S4 slicer uses the path from any point (here, Benchy’s prow) as its basis…
This non-planar slicer is built into a Jupyter notebook, which follows a relatively simple algorithm to automatically generate non-planar toolpaths for any model. It does this by first generating a tetrahedral mesh of the model and then calculating the shortest possible path through the model from any given tetrahedron to the print bed. Even with non-planar printing, you need to print from the print-bed up (or out).

Quite a lot of math is done to use these paths to calculate a deformation mesh, and we’ll leave that to [Joshua] to explain in his video below. After applying the deformation, he slices the resulting mesh in Cura, before the G-code goes back to Jupyter to be re-transformed, restoring the shape of the original mesh.
… to generate deformed models for slicing, like this.
So yes, it is G-code bending as others have demonstrated before, but in a reproducible, streamlined, and straightforward workflow. Indeed, [Josh] credits much of the work to earlier work on the S^3-Slicer, which inspired much of the logic and the name behind his S4 slicer. (Not S4 as in “more than S^3” but S4 as a contraction of “Simplified S^3”). Once again, open source allows for incremental innovation.

It is admittedly a computationally intensive process, and [Joshua] uses a simplified model of Benchy for this demo. This seems exactly the sort of thing we’d like to burn compute power on, though.

This sort of non-planar 3D printing is an exciting frontier, one which we have covered before. We’ve seen techniques for non-planar infill, or even to print overhangs on unmodified Cartesian printers, but this is probably the first time we’ve seen Benchy given the non-planar treatment. You can try S4 slicer for yourself via GitHub, or just watch the non-planar magic in action after the break.

youtube.com/embed/M51bMMVWbC8?…

hackaday.com/2025/04/20/non-pl…

L’algoritmo Dreamer AI, sviluppato dai ricercatori di Google DeepMind in collaborazione con uno scienziato dell’Università di Toronto, ha imparato a migliorare le proprie abilità autonomamente, padroneggiando Minecraft in pochi giorni. Come riportato nell’ultimo numero di Nature, Dreamer ha raggiunto il livello esperto in soli nove giorni, senza alcuna formazione precedente: non gli è stato spiegato come giocare, ma gli è stato semplicemente fornito un quadro di possibili azioni e un sistema di ricompense per chi trova i diamanti.

Sebbene addestrare l’intelligenza artificiale tramite giochi possa sembrare uno spreco di risorse (è solo un computer che gioca con un altro computer), in realtà fornisce informazioni su come far sì che l’intelligenza artificiale si adatti, impari e raggiunga obiettivi in ​​ambienti mutevoli. E questo è già un passo importante verso robot in grado di svolgere compiti nel mondo reale senza supervisione e intervento costanti.

Durante l’esperimento, Dreamer è stato inserito in un ambiente di gioco Minecraft, in cui ogni sessione di gioco durava solo 30 minuti. Dopodiché il gioco è ripreso in un mondo virtuale completamente nuovo, con paesaggi, risorse e condizioni diverse.

Nonostante queste limitazioni, l’algoritmo ha imparato rapidamente: ha immaginato un futuro in cui aveva già raggiunto tutti i suoi obiettivi e si è concentrato solo sulle azioni che avrebbero portato a quel risultato. Queste abilità vennero mantenute e utilizzate nel successivo mondo virtuale, aiutando Dreamer a migliorare rapidamente le sue capacità.

Il punto è che Dreamer non impara copiando la strategia di qualcun altro, ma sperimentando e tracciando attentamente un percorso verso il successo. Ciò apre importanti possibilità: se l’intelligenza artificiale è in grado di raggiungere autonomamente obiettivi complessi in tali condizioni, può potenzialmente imparare a risolvere problemi reali nel mondo fisico, dalla navigazione alla riparazione o manutenzione.

E forse molto presto i robot saranno effettivamente in grado di imparare da soli, senza dover attendere istruzioni dagli esseri umani.

L'articolo Dreamer AI domina Minecraft in 9 giorni: l’algoritmo che impara da solo è arrivato proviene da il blog della sicurezza informatica.

La società di sicurezza Prodaft ha annunciato il lancio dell’iniziativa SYS (Sell your Source) , nell’ambito della quale è pronta ad acquistare vecchi account su noti forum di hacker e siti darknet. Promettono di pagare di più per gli account dei moderatori e degli amministratori.

“Se a un certo punto della tua vita hai creato un account su un forum darknet che non faceva parte dei tuoi abituali social media, e se sei pronto a lasciarti alle spalle quel passato, allora questo messaggio è per te”, scrive l’azienda.

“Forse è successo nel 2015, quando ti sei immerso in argomenti che erano troppo rischiosi per Google? O nel 2020, quando hai provato ad acquistare online cose che ti sembravano sospette o pericolose, come sostanze che promettevano di curare un virus? Non ha molta importanza cosa stessi facendo in quel periodo. Ciò che conta adesso è se vuoi lasciare questa parte nel passato.”

I rappresentanti di Prodaft spiegano che gli account verificati e vecchi sui siti underground li aiuteranno nel loro lavoro, poiché gli account sui forum degli hacker consentono loro sostanzialmente di osservare il panorama delle minacce dall’interno.

Attualmente, gli account su XSS, Exploit[.]in, RAMP4U, Verified e BreachForums registrati prima di dicembre 2022 vengono presi in considerazione per l’acquisto. Promettono di pagare di più per gli account con diritti di moderatore e amministratore.

L’azienda assicura che non ci saranno giudizi né domande inutili, “una transazione semplice e sicura che avvantaggia entrambe le parti”. Si sottolinea che tutti i dati sugli account acquistati saranno trasferiti alle forze dell’ordine, ma la società promette di “proteggere in modo sicuro l’identità del venditore” e di non divulgare ulteriori informazioni alle autorità.

L'articolo Prodaft compra account su forum underground. XSS, Exploit, RAMP4U e 5.000€ per ex moderatori proviene da il blog della sicurezza informatica.