Remember the coolest geometry set you ever lay eyes on as a school-going kid? It could be something you owned, saw someone else own, or just glanced wishfully at while at a store. Now imagine all that coolness packed not into a set, but into one singular instrument. When I saw the Tiruler for the first time, thats pretty much the feeling I got that of awe, euphoria, and sheer desire. Weve all seen folding rulers before, but imagine one that has a caliper, a pencil, a screw-driver, or a spirit level too. Sounds wild, doesnt it? Now imagine the entire thing made out of titanium, so even though you buy it out of being an excited EDC enthusiast, you can pass it down like a family heirloom to your kids when they need a cool geometry set of their own.Born from the engineering minds at MakerPi, the Tiruler represents a fascinating pivot into the tangible, analog world of everyday carry. With more than 10 functions built into its unique form, the Tiruler is like nothing Ive personally ever seen. Every single inch of the gizmo hides a feature that gives the Tiruler incredible meaning and value all blended into an industrial-steampunk aesthetic that isnt merely a visual flourish the exposed mechanics and dual-material construction telegraph the tools serious intentions while celebrating its utilitarian soul.Designer: MakerPiClick Here to Buy Now: $89 $138 (39% off) Hurry! Only 6 days left.The best way to appreciate the Tiruler is to really look at all that it packs into its compact form. The device measures roughly 120mm or 4.7 inches when closed, opening up into a 200mm or 7.8-inch scale, depending on your need. The entire things made from titanium with etched numbers, so you could use it to measure, draw, cut, or do anything without worrying about any wear or tear. The pivoting scale expands based on need, but heres the kicker, it also packs both imperial and metric measurements, and that pivot? It works as a protractor.The pivot features a brass enclosure with engraved angles that let you use the Tiruler to draw or measure angles. In all fairness, I had a scale/ruler that did this in school but it was made out of plastic, had its fair share of errors, and wasnt as immaculately produced as this. Here, titanium-machined precision makes the Tiruler fantastically accurate. Want me to elaborate? The thing has a literal vernier caliper built into it too! But well get to that later on.Along the pivoting mechanism is also a rolling scale that lets you measure non-linear surfaces. Perfect for curves, or just measuring something you cant hold a scale to, the rolling scale lets you roll along the line you want to measure. Every time it completes a revolution, it clicks, making keeping track of measurements easy. This effectively means you can go on measuring indefinitely just count the clicks and then look at the final number on the roller to get your overall calculation.For more complex measurements, an integrated vernier caliper gives you the ability to do stuff your regular ruler cant. The calipers jaw sits inside a perfectly machined slot on the Tiruler, sliding out as much as 50mm, while giving you measurements of 0.1mm of accuracy, or 0.003 inches. A button on the end makes sliding easy, unlocking the calipers jaw just when you need it and not otherwise.When youre not measuring, youre probably drawing which is why the Tiruler packs its own Eternal Pencil into its design. Sitting snugly in a slot thanks to magnets, the pencil can be undocked whenever needed, and used to draw lines, make doodles, annotate documents with measurements, etc. The eternal pencil tip never needs sharpening and can work for decades, while the pencils metal body also packs a sharp-ish end thats perfect for running along the edge of boxes to open them without really being a traditional blade.Other rather clever details include the presence of a hex slot for mounting hex bits that turn the Tiruler into a high-torque screwdriver, a spirit level that lets you use the Tiruler to align objects like, say, paintings on a wall, and the mark of a good EDC a bottle opener that lets you hang loose between serious measuring jobs. Maybe not the best feature for a ruler youd pass down to your kid, but hey, the Tiruler is versatile EDC first and heirloom-material later. Finally, tritium slots really cement the EDC status of the Tiruler, making it the perfect collector-item for anyone who loves gear and multitools.The choice of titanium for the Tirulers construction gives it the power of true longevity. Beyond its obvious resistance to bending (the death knell for lesser rulers), titanium offers corrosion resistance that ensures this tool will age gracefully through years of workshop abuse and pocket carry. The overall device weighs a paltry 60 grams or 2.1 ounces, and the lack of a blade makes the Tiruler TSA-friendly, so you can easily carry it with you wherever you go whether its to work or even on a flight. Now I dont normally need a ruler or protractor on flights, but hey, you do you?What impresses most isnt just the number of functions MakerPi has packed into the Tiruler, but how harmoniously they coexist. Each capability occupies its dedicated space without encroaching on the others, resulting in a tool that performs each task with surprising competence. The Tiruler starts at $89 (surprisingly affordable for a titanium-based multitool), which includes the ruler itself, the eternal pencil, and also a cowhide leather holder for your tool, giving you something thats truly upper-echelon. I swear, if I had this as a child Id be the coolest kid in math class.Click Here to Buy Now: $89 $138 (39% off) Hurry! Only 6 days left.The post This Steampunk Titanium EDC Ruler Packs 10 Surprising Tools You Didnt Know You Needed first appeared on Yanko Design.

We found the best protein powders that won't make your shake taste like drywall.

The use of a video persona created with artificial intelligence software to help make an argument earns a stern rebuke.

One of Apple's most controversial stores never opened, but a virtual tour that can be viewed on Apple Vision Pro brings the dream back to life.Apple's Federation Square building render. Credit: Filip ChudzinskiApple's plan to build a flagship store in Melbourne's Federation Square sparked immediate backlash and was ultimately canceled. Now, a virtual recreation offers a glimpse of what could have been.In 2017, Apple announced its intention to open a global flagship in the heart of Melbourne, marking its first in the Southern Hemisphere. Designed in partnership with Foster + Partners, the multilevel glass pavilion would have anchored the southeast corner of Federation Square. Continue Reading on AppleInsider | Discuss on our Forums

Researchers from Northwestern University have developed a carbon-negative building material that is made using seawater, electricity, and carbon dioxide. Their approach focuses on creating an alternative source of the calcium- and magnesium-based minerals that comprise materials, such as cement, concrete, paint, and plasters.The material is created by inserting electrodes into seawater and applying an electric current. Carbon dioxide is then bubbled into the water, which changes its chemical composition resulting in the creation of solid minerals, including calcium carbonate and magnesium hydroxide. The lead of the study, Professor Alessandro Rotta Loria, compares the process to the way coral and mollusks form their shells.The material can reportedly hold over half of its weight in carbon dioxide, depending on the ratio of the minerals. Loria also says the material would not weaken the strength of concrete or cement.

Kirby Air Riderswas one of several first-party titles that Nintendo announced for the Switch 2 during its recent Direct presentation, though it revealed little about the game with a brief CG teaser. Wedid,however, get confirmation that Masahiro Sakurai theSuper Smash Bros.mastermind who also directed 2003sKirby Air Ride is returning 22 years on to direct the sequel. Now, Sakurai himself has revealed another crucial tidbit.Taking to Twitter, Sakurai recently revealed thatKirby Air Ridersis going to be developed by Bandai Namco Studios. Of course, this isnt the first time that Bandai Namco will develop a Masahiro Sakurai-directed title, having led development onSuper Smash Bros. for Wii Uand3DS, as well asSuper Smash Bros. Ultimate.Bandai Namco has also worked on other first-party Nintendo titles in the form ofMario Kart 7, Mario Kat Tour, Mario Sports Superstars, andARMS.In 2023, Bandai Namco formally announced that it had established Studio 2 and Studio S, a team dedicated to working on projects commissioned by Nintendo.Kirby Air Riderswill launch exclusively for the Nintendo Switch 2 sometime this year.Just like Super Smash Bros. for Wii U & 3DS and Ultimate, the recently announced Kirby Air Riders game is also being developed by Bandai Namco Studios.The full staff is giving their all in developing the game, but it will take some time before we can share more information. / Masahiro Sakurai (@Sora_Sakurai) April 3, 2025

Cool FindsRare Early Beatles Demo Tape Found Collecting Dust in Vancouver Record StoreThe 15-song recording dates to the Liverpool bands failed audition for Decca Records in early 1962months before it released its first hit Before Beatlemania, the Liverpool band was rejected from Decca Records in 1962. Public domain via Wikimedia CommonsOn New Years Day in 1962, a four-man band from Liverpool arrived at Decca Records in London to record a demo tape. The bands 15-songaudition was apparently underwhelming:The Beatles were rejected from the label in favor of a group called Brian Poole and the Tremeloes.The rejection was stinging for theyoung band, but it was soon forgotten. By the spring, the Beatles hadsigned with EMIs Parlophone Records. Later that summer, the band solidified its lineup, replacing drummer Pete Best with Ringo Starr. Its debut single, Love Me Do, came out in October, and stardom soon followed.The audition tapes, too, were soon forgotten, aside from some incomplete bootlegs that circulated among die-hard fans and five songs released on the 1995 compilation album Anthology 1.So when Rob Frith, the owner ofNeptoon Records in Vancouver, brought a tape labeled Beatles 60s demos to his friends studio on a whim, he didnt know the value of what he possessed.Although the tape had been collecting dust behind his counter for years, he had never given it a listen. I thought it was just a reel-to-reel tape that somebody had put bootleg things on, he tells Tessa Vikander of the Canadian Broadcasting Corporation (CBC).When Friths disc-jockey friend Larry Hennessey played the recording on his vintageStuder A810 tape player last month, however, they knew they had something special. The first thing that struck them was the white leader tape on the record, an extra measure of care to create separation between songswhich would have been atypical for a bootleg. Then, the quality of the recording blew them away.Right away, were all kind of looking at each other, Frith tells theNew York Times Neil Vigdor. It seems like the Beatles are in the room. Thats how clear it is.Doug Schober, another friend and music aficionado in the room at the time, began Googling around to find out what he could about the tape. By the third song, I say, I think this is the Decca demo, he tells the Times.Frith posted ashort clip of the tape on social media, writing, After hearing it last night for the first time, it sounds like a master tape.Indeed, it was a direct copy of the master tape of the infamous Decca audition, with all 15 songs, including three originals, in crystal-clear quality. With the help of social media, Frith and his friends were able to track down the provenance of this rare find.Replying to Friths Facebook post, Tom Lavin, a member of the Powder Blues Band, recalled that the tape used to be at Vancouvers Can-Base Studios after its owner, Jack Herschorn, obtained it in London from someone inside Decca, per theVancouver Suns John Mackie.Herschorn, who now lives in Mexico, confirmed Lavins story. He claims that a well-known record producer gave him the tape in 1968 or 1969 hoping that he would circulate it as a bootleg in North America, according to the Sun.But Herschorn demurred. I wouldnt want somebody doing that to me, he tells the Sun. It was just a moral issue with me. I could have put it out, made a few bucks on it, but then I could get bad PR get sued over. It wasnt my style.The Sheik Of Araby (Anthology 1 Version)Watch on Thats how the tape ended up in storage at Can-Base studios, where musicians like Lavin listened to it for inspiration without ever realizing the tapes true significance.When Herschorn moved and sold the studio, he left the tape behind. I should have [taken] it, but it didnt work out that way, he explains to the CBC.Whoever took the demo tape from the studios may not have realized what they had when they sold it to Frith. I actually cant remember who I bought it from, Frith tells the Sun. He surely didnt know that he was purchasing a piece of music history.The Beatles, for that matter, didnt think much of the audition tape, either.Listening to the tapes, I can understand why we failed the Decca audition,Paul McCartney said inThe Beatles Anthology, an oral history published in 2000. We werent that good; though there were some quite interesting and original things.The three original songs the Beatles tried out for Decca were Like Dreamers Do, Hello Little Girl and Love of the Loved. The band also played standards like Money (Thats What I Want) that showed up on later studio albums.I think Decca expected us to be all polished,John Lennon reflected in an interview reprinted in Anthology. We were just doing a demo. They should have seen our potential.In retrospect,George Harrison found the matter altogether amusing. Years later, he learned of the canny prediction that the head of Decca records had given Brian Epstein, the Beatles manager: Guitar groups are on the way out.Clearly, the Decca tapes signaled anything but the end of guitar groups or the Beatles music career.In 2012, a tape of the Decca auditionsold at auction for 35,000 (more than $56,000), but it only contained ten songs. No full 15-song tape with such high fidelity has ever surfaced before.Frith is planning to keep the one-of-a-kind tape for now, although he tells the Times he would consider giving it to McCartney for an official release or holding a listening event for charity.A month ago, Frith adds, if someone had given me 20 bucks for that tape, I probably would have sold it.Get the latest stories in your inbox every weekday.

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn MoreWeaponized large language models (LLMs) fine-tuned with offensive tradecraft are reshaping cyberattacks, forcing CISOs to rewrite their playbooks. Theyve proven capable of automating reconnaissance, impersonating identities and evading real-time detection, accelerating large-scale social engineering attacks.Models, includingFraudGPT,GhostGPT andDarkGPT, retail for as little as $75 a month andare purpose-built for attack strategies such as phishing, exploit generation, code obfuscation, vulnerability scanning and credit card validation.Cybercrime gangs, syndicates and nation-states see revenue opportunities in providing platforms, kits and leasing access to weaponized LLMs today. These LLMs are being packaged much like legitimate businesses package and sell SaaS apps. Leasing a weaponized LLM often includes access to dashboards, APIs, regular updates and, for some, customer support.VentureBeat continues to track the progression of weaponized LLMs closely. Its becoming evident that the lines are blurring between developer platforms and cybercrime kits as weaponized LLMs sophistication continues to accelerate. With lease or rental prices plummeting, more attackers are experimenting with platforms and kits, leading to a new era of AI-driven threats.Legitimate LLMs in the cross-hairsThe spread of weaponized LLMs has progressed so quickly that legitimate LLMs are at risk of being compromised and integrated into cybercriminal tool chains. The bottom line is that legitimate LLMs and models are now in the blast radius of any attack.The more fine-tuned a given LLM is, the greater the probability it can be directed to produce harmful outputs. Ciscos The State of AI Security Report reports that fine-tuned LLMs are 22 times more likely to produce harmful outputs than base models. Fine-tuning models is essential for ensuring their contextual relevance. The trouble is that fine-tuning also weakens guardrails and opens the door to jailbreaks, prompt injections and model inversion.Ciscos study proves that the more production-ready a model becomes, the more exposed it is to vulnerabilities that must be considered in an attacks blast radius. The core tasks teams rely on to fine-tune LLMs, including continuous fine-tuning, third-party integration, coding and testing, and agentic orchestration, create new opportunities for attackers to compromise LLMs.Once inside an LLM, attackers work fast to poison data, attempt to hijack infrastructure, modify and misdirect agent behavior and extract training data at scale. Ciscos study infers that without independent security layers, the models teams work so diligently on to fine-tune arent just at risk; theyre quickly becoming liabilities. From an attackers perspective, theyre assets ready to be infiltrated and turned.Fine-Tuning LLMs dismantles safety controls at scaleA key part of Ciscos security teams research centered on testing multiple fine-tuned models, including Llama-2-7B and domain-specialized Microsoft Adapt LLMs. These models were tested across a wide variety of domains including healthcare, finance and law.One of the most valuable takeaways from Ciscos study of AI security is that fine-tuning destabilizes alignment, even when trained on clean datasets. Alignment breakdown was the most severe in biomedical and legal domains, two industries known for being among the most stringent regarding compliance, legal transparency and patient safety.While the intent behind fine-tuning is improved task performance, the side effect is systemic degradation of built-in safety controls. Jailbreak attempts that routinely failed against foundation models succeeded at dramatically higher rates against fine-tuned variants, especially in sensitive domains governed by strict compliance frameworks.The results are sobering. Jailbreak success rates tripled and malicious output generation soared by 2,200% compared to foundation models. Figure 1 shows just how stark that shift is. Fine-tuning boosts a models utility but comes at a cost, which is a substantially broader attack surface.TAP achieves up to 98% jailbreak success, outperforming other methods across open- and closed-source LLMs. Source: Cisco State of AI Security 2025, p. 16.Malicious LLMs are a $75 commodityCisco Talos is actively tracking the rise of black-market LLMs and provides insights into their research in the report. Talos found that GhostGPT, DarkGPT and FraudGPT are sold on Telegram and the dark web for as little as $75/month. These tools are plug-and-play for phishing, exploit development, credit card validation and obfuscation.DarkGPT underground dashboard offers uncensored intelligence and subscription-based access for as little as 0.0098 BTCframing malicious LLMs as consumer-grade SaaS.Source: Cisco State of AI Security 2025, p. 9.Unlike mainstream models with built-in safety features, these LLMs are pre-configured for offensive operations and offer APIs, updates, and dashboards that are indistinguishable from commercial SaaS products.$60 dataset poisoning threatens AI supply chainsFor just $60, attackers can poison the foundation of AI modelsno zero-day required, write Cisco researchers. Thats the takeaway from Ciscos joint research with Google, ETH Zurich and Nvidia, which shows how easily adversaries can inject malicious data into the worlds most widely used open-source training sets.By exploiting expired domains or timing Wikipedia edits during dataset archiving, attackers can poison as little as 0.01% of datasets like LAION-400M or COYO-700M and still influence downstream LLMs in meaningful ways.The two methods mentioned in the study, split-view poisoning and frontrunning attacks, are designed to leverage the fragile trust model of web-crawled data. With most enterprise LLMs built on open data, these attacks scale quietly and persist deep into inference pipelines.One of the most startling discoveries Cisco researchers demonstrated is that LLMs can be manipulated to leak sensitive training data without ever triggering guardrails. Cisco researchers used a method called decomposition prompting to reconstruct over 20% of select New York Times and Wall Street Journal articles. Their attack strategy broke down prompts into sub-queries that guardrails classified as safe, then reassembled the outputs to recreate paywalled or copyrighted content.Successfully evading guardrails to access proprietary datasets or licensed content is an attack vector every enterprise is grappling to protect today. For those that have LLMs trained on proprietary datasets or licensed content, decomposition attacks can be particularly devastating. Cisco explains that the breach isnt happening at the input level, its emerging from the models outputs. That makes it far more challenging to detect, audit or contain.If youre deploying LLMs in regulated sectors like healthcare, finance or legal, youre not just staring down GDPR, HIPAA or CCPA violations. Youre dealing with an entirely new class of compliance risk, where even legally sourced data can get exposed through inference, and the penalties are just the beginning.Ciscos ongoing research, including Talos dark web monitoring, confirms what many security leaders already suspect: weaponized LLMs are growing in sophistication while a price and packaging war is breaking out on the dark web. Ciscos findings also prove LLMs arent on the edge of the enterprise; they are the enterprise. From fine-tuning risks to dataset poisoning and model output leaks, attackers treat LLMs like infrastructure, not apps.One of the most valuable key takeaways from Ciscos report is that static guardrails will no longer cut it. CISOs and security leaders need real-time visibility across the entire IT estate, stronger adversarial testing, and a more streamlined tech stack to keep up and a new recognition that LLMs and models are an attack surface that becomes more vulnerable with greater fine-tuning.Daily insights on business use cases with VB DailyIf you want to impress your boss, VB Daily has you covered. We give you the inside scoop on what companies are doing with generative AI, from regulatory shifts to practical deployments, so you can share insights for maximum ROI.Read our Privacy PolicyThanks for subscribing. Check out more VB newsletters here.An error occured.

Earlier this week, when it seemed as though TikToks fate in the US would actually be decided by April 5th, everyone from Amazon to the founder of OnlyFans was coming out of the woodwork to buy it.As it turns out, none of them had a chance. And now, thanks to President Donald Trumps tariff war, no one may get to buy TikTok. People familiar with the matter tell me that, despite all of the bids for the app, the White House was only seriously considering an Oracle-led consortium, which included many of ByteDances biggest investors who were set to roll their stakes into a new, US entity. The proposal, which would have licensed the apps algorithm from China and shuffled some shareholder money around to make TikTok look more independent from ByteDance, was set to be announced before President Trump went nuclear on tariffs. As others have reported and Ive independently confirmed, his tariff announcement on Wednesday torched any immediate chance of the TikTok proposal being blessed by the Chinese government. On Friday, less than an hour after Trump said he was pushing back the clock on banning TikTok by another 75 days to finish working out a deal, ByteDance issued its first statement on the situation, saying that any agreement will be subject to approval under Chinese law. After suggesting on Thursday that China would approve a TikTok deal in exchange for tariff relief, the president seemed less confident on Friday: We hope to continue working in Good Faith with China, who I understand are not very happy about our Reciprocal Tariffs (Necessary for Fair and Balanced Trade between China and the U.S.A.!).Anyone who has been closely following this tortured saga that is the TikTok ban is exhausted. Now, five years after he first tried to ban it, Trump has seemingly torpedoed the plan that was finally going to keep TikTok running. His friend Larry Ellison was going to secure more business for Oracle by reviving the broad strokes of the Project Texas security proposal that was made to the previous administration (albeit this time with fewer security guarantees, since this administration doesnt really care about national security threats the app may or may not pose). TikToks investors and employees were finally going to get certainty about the apps fate. ByteDance was going to effectively stay in the drivers seat while maintaining control of its algorithm. And Trump was going to get to say that he saved TikTok. TikTok gets to continue operating in the US for a while longer, thanks to Trumps Department of Justice not enforcing a ban that was passed by Congress and upheld by the Supreme Court. TikToks US hosting providers, including Apple, Google, and Oracle, are seemingly still okay with this situation despite the fact that they are very much breaking the law. And unless Trump ultimately eases on tariffs against China, TikTok may end up banned after all.What a giant farce. OpenAI is going to release its o3 and o4-mini models in a couple of weeks, and then do GPT-5 in a few months.DOGE staffers have arrived at the FTC.Microsofts 50th anniversary celebration was disrupted by multiple employees protesting its AI business with the Israeli government.Automattic, the parent company of WordPress, is laying off about 16 percent of employees.Adobe is the latest company to axe DEI hiring goals.Job boardSome noteworthy job changes in the tech world:Joelle Pineaus departure from Meta means that Yann LeCun is the interim head of AI research while the company looks for a permanent leader. When I last interviewed her in October, Pineau seemed very bought into her role at Meta. Since then, Im told she has privately expressed concern with the companys Trump-fueled policy changes, which Im told are also impacting the upcoming Llama 4 release. More broadly, her exit signals that, like its peers, Meta is shifting its AI research from being more free-wheeling towards directly supporting its business goals. Well, thats awkward! Less than two weeks after she appeared as the main character in a glossy magazine story about Googles efforts to beat ChatGPT, Gemini product chief Sissie Hsiao is out (though, in classic Google fashion, apparently finding a role in the company elsewhere). The Gemini app teams new leader is Josh Woodward, who also leads the Labs org behind the breakout hit that has been NotebookLM. While Woodward seems to be generally well-liked inside Google, Hsiaos reign has been, to put it mildly, controversial. Sundar Pichais main goal is for the Gemini app to surpass ChatGPT this year. The pressure is on.In the wake of co-CEO Han Jong-Hees death, Samsungs mobile business chief, TM Roh, is taking on an expanded role also overseeing the broader devices group.Ring founder Jamie Siminoff is back at Amazon and also leading its Blink, Amazon Key, and Sidewalk teams. Deels head of communications, Elisabeth Diana, resigned amid the HR startups spy scandal with Rippling. Tim Zaman, a former senior AI infrastructure engineer for Elon Musk, is joining OpenAI after a stint at DeepMind to work on frontier clusters. More linksPolitico looks at how the tech world is scared of Elon Musk.Photos of the mansion Mark Zuckerberg just bought in Washington, DC.Tech companies are advising employees on visas not to leave the U.S. Bloomberg charts the new billionaires of the AI boom.Keach Hageys forthcoming book about OpenAI has a juicy chapter about how Sam Altman was fired by the board.The AI industry is arguing about this timeline prediction for how the technology will play out.Alex Konrad interviews the CEO of Granola, one of the few AI apps I find to be actually useful.Our hands-on with the Nintendo Switch 2.If you havent already, dont forget to subscribe to The Verge, which includes unlimited access to Command Line and all of our reporting.As always, I want to hear from you, especially if you have feedback on this issue or a story tip. Respond here or ping me securely on Signal.Thanks for subscribing.See More:

In todays information-rich digital landscape, navigating extensive web content can be overwhelming. Whether youre researching for a project, studying complex material, or trying to extract specific information from lengthy articles, the process can be time-consuming and inefficient. This is where an AI-powered Question-Answering (Q&A) bot becomes invaluable.This tutorial will guide you through building a practical AI Q&A system that can analyze webpage content and answer specific questions. Instead of relying on expensive API services, well utilize open-source models from Hugging Face to create a solution thats:Completely free to useRuns in Google Colab (no local setup required)Customizable to your specific needsBuilt on cutting-edge NLP technologyBy the end of this tutorial, youll have a functional web Q&A system that can help you extract insights from online content more efficiently.What Well BuildWell create a system that:Takes a URL as inputExtracts and processes the webpage contentAccepts natural language questions about the contentProvides accurate, contextual answers based on the webpagePrerequisitesA Google account to access Google ColabBasic understanding of PythonNo prior machine learning knowledge requiredStep 1: Setting Up the EnvironmentFirst, lets create a new Google Colab notebook. Go to Google Colab and create a new notebook.Lets start by installing the necessary libraries:# Install required packages!pip install transformers torch beautifulsoup4 requestsThis installs:transformers: Hugging Faces library for state-of-the-art NLP modelstorch: PyTorch deep learning frameworkbeautifulsoup4: For parsing HTML and extracting web contentrequests: For making HTTP requests to webpagesStep 2: Import Libraries and Set Up Basic FunctionsNow lets import all the necessary libraries and define some helper functions:import torchfrom transformers import AutoModelForQuestionAnswering, AutoTokenizerimport requestsfrom bs4 import BeautifulSoupimport reimport textwrap# Check if GPU is availabledevice = torch.device('cuda' if torch.cuda.is_available() else 'cpu')print(f"Using device: {device}")# Function to extract text from a webpagedef extract_text_from_url(url): try: headers = { 'User-Agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/91.0.4472.124 Safari/537.36' } response = requests.get(url, headers=headers) response.raise_for_status() soup = BeautifulSoup(response.text, 'html.parser') for script_or_style in soup(['script', 'style', 'header', 'footer', 'nav']): script_or_style.decompose() text = soup.get_text() lines = (line.strip() for line in text.splitlines()) chunks = (phrase.strip() for line in lines for phrase in line.split(" ")) text = 'n'.join(chunk for chunk in chunks if chunk) text = re.sub(r's+', ' ', text).strip() return text except Exception as e: print(f"Error extracting text from URL: {e}") return NoneThis code:Imports all necessary librariesSets up our device (GPU if available, otherwise CPU)Creates a function to extract readable text content from a webpage URLStep 3: Load the Question-Answering ModelNow lets load a pre-trained question-answering model from Hugging Face:# Load pre-trained model and tokenizermodel_name = "deepset/roberta-base-squad2"print(f"Loading model: {model_name}")tokenizer = AutoTokenizer.from_pretrained(model_name)model = AutoModelForQuestionAnswering.from_pretrained(model_name).to(device)print("Model loaded successfully!")Were using deepset/roberta-base-squad2, which is:Based on RoBERTa architecture (a robustly optimized BERT approach)Fine-tuned on SQuAD 2.0 (Stanford Question Answering Dataset)A good balance between accuracy and speed for our taskStep 4: Implement the Question-Answering FunctionNow, lets implement the core functionality the ability to answer questions based on the extracted webpage content:def answer_question(question, context, max_length=512): max_chunk_size = max_length - len(tokenizer.encode(question)) - 5 all_answers = [] for i in range(0, len(context), max_chunk_size): chunk = context[i:i + max_chunk_size] inputs = tokenizer( question, chunk, add_special_tokens=True, return_tensors="pt", max_length=max_length, truncation=True ).to(device) with torch.no_grad(): outputs = model(**inputs) answer_start = torch.argmax(outputs.start_logits) answer_end = torch.argmax(outputs.end_logits) start_score = outputs.start_logits[0][answer_start].item() end_score = outputs.end_logits[0][answer_end].item() score = start_score + end_score input_ids = inputs.input_ids.tolist()[0] tokens = tokenizer.convert_ids_to_tokens(input_ids) answer = tokenizer.convert_tokens_to_string(tokens[answer_start:answer_end+1]) answer = answer.replace("[CLS]", "").replace("[SEP]", "").strip() if answer and len(answer) > 2: all_answers.append((answer, score)) if all_answers: all_answers.sort(key=lambda x: x[1], reverse=True) return all_answers[0][0] else: return "I couldn't find an answer in the provided content."This function:Takes a question and the webpage content as inputHandles long content by processing it in chunksUses the model to predict the answer span (start and end positions)Processes multiple chunks and returns the answer with the highest confidence scoreStep 5: Testing and ExamplesLets test our system with some examples. Heres the complete code:url = "https://en.wikipedia.org/wiki/Artificial_intelligence"webpage_text = extract_text_from_url(url)print("Sample of extracted text:")print(webpage_text[:500] + "...")questions = [ "When was the term artificial intelligence first used?", "What are the main goals of AI research?", "What ethical concerns are associated with AI?"]for question in questions: print(f"nQuestion: {question}") answer = answer_question(question, webpage_text) print(f"Answer: {answer}")This will demonstrate how the system works with real examples.Output of the above codeLimitations and Future ImprovementsOur current implementation has some limitations:It can struggle with very long webpages due to context length limitationsThe model may not understand complex or ambiguous questionsIt works best with factual content rather than opinions or subjective materialFuture improvements could include:Implementing semantic search to better handle long documentsAdding document summarization capabilitiesSupporting multiple languagesImplementing memory of previous questions and answersFine-tuning the model on specific domains (e.g., medical, legal, technical)ConclusionNow youve successfully built your AI-powered Q&A system for webpages using open-source models. This tool can help you:Extract specific information from lengthy articlesResearch more efficientlyGet quick answers from complex documentsBy utilizing Hugging Faces powerful models and the flexibility of Google Colab, youve created a practical application that demonstrates the capabilities of modern NLP. Feel free to customize and extend this project to meet your specific needs.Useful ResourcesHere is the Colab Notebook. Also,dont forget to follow us onTwitterand join ourTelegram ChannelandLinkedIn Group. Dont Forget to join our85k+ ML SubReddit. Mohammad AsjadAsjad is an intern consultant at Marktechpost. He is persuing B.Tech in mechanical engineering at the Indian Institute of Technology, Kharagpur. Asjad is a Machine learning and deep learning enthusiast who is always researching the applications of machine learning in healthcare.Mohammad Asjadhttps://www.marktechpost.com/author/mohammad_asjad/DeltaProduct: An AI Method that Balances Expressivity and Efficiency of the Recurrence Computation, Improving State-Tracking in Linear Recurrent Neural NetworksMohammad Asjadhttps://www.marktechpost.com/author/mohammad_asjad/PydanticAI: Advancing Generative AI Agent Development through Intelligent Framework DesignMohammad Asjadhttps://www.marktechpost.com/author/mohammad_asjad/TxAgent: An AI Agent that Delivers Evidence-Grounded Treatment Recommendations by Combining Multi-Step Reasoning with Real-Time Biomedical Tool IntegrationMohammad Asjadhttps://www.marktechpost.com/author/mohammad_asjad/Building a Retrieval-Augmented Generation (RAG) System with FAISS and Open-Source LLMs