Goodbye GPT-5... New DeepSeek Update is HERE! AI News EXPLAINED

Deepseek just rolled out a fresh model and it straight up smokes Chad GBC faster, slicker, and still totally free. The new drop is wild. So, if you've been sleeping on it, now is the time to grab the model and spin it up. I bet some Alman squad is already poking around the code trying to snag a few tricks before GBT ships. While they're busy digging for secrets, let me give you the lowdown on Deep Seek's upgrade and a stack of other AI news you don't want to miss.

So, Deep Seek just dropped a new model R1052A. Yeah, the name sounds like a Star Wars droid nobody asked for, but whatever. This thing hits hard. They claim it reasons deeper than anything they've ever shipped. I've been poking it all morning, and no lie, it's sharper. On their own chart, the line practically hugs OpenAI 3. And in some tests, it even nose has passed Google's Gemini 2. 5 Pro. They cranked the training on logic, puzzles, math, and coding. Plus, they finally said the big scary word, hallucinations. Props for the honesty. Most teams hide behind a improved accuracy. Like, we can see the emperor's new clothes. Hallucinations, by the way, are still the big clown in the room. One second, your AI's genius. Next second, it's given friends a brand new capital city. At least Deep Seek's calling it out instead of sweeping it under the marketing rug. Interface nerds, don't get excited. Same old layout, same buttons. The magic is under the hood. Web search finally behaves. I used to get something went wrong every other click. Now it's smooth. Life effects pop in without the model losing the plot. It still has fewer toys than Chad GBT, but raw brain power feels scarily close. And as a bonus, you can download the model, the hall model, fire it up in a halfway decent laptop while your coffee brews and boom, private LM at your fingertips. Is it perfect? Too early to crown it. Real tests come when we throw our messy everyday stuff at it, but speedwise it's neck andneck with GBT3. Last year that sounded like sci-fi. Now a scrappy crew with couch cushion money is nipping at OpenAI's heels. Wild times, man. Can't wait to see where we are next year. The new Deep Seek model is undoubtedly cool, but I wonder if they finally fixed its web search. It didn't work before, and I think it's because of the way Deepseek was scraping the web for data. When you do any kind of automation, sites usually fight back with smart detection tools. If lots of requests come from the same IP address, the site blocks you. That's what might have been a problem. The simple fix is to use clean, highquality proxies. All Deepseek needed to do is get a subscription to Node Maven, the sponsor of today's video. Over 90% of their IPs are clean and safe, while the biggest brands could only guarantee around 25%. Fewer bad IPs means fewer bands, fewer captures, and a lot less stress. And this quality isn't just words. You can check it. There are many extensions for your browser that do that, like pixel scan proxy checker. So, you can go to Node Maven and generate yourself a ton of configurable and customizable proxies. Then, you can copy them and paste them into the extension and it will say that the quality is high in all of them. And like I said, all proxies are customizable. Choose your own location, session type, or even proxy access. And all that comes with Node Maven's industry first filtration pipeline. No address reaches a customer before it has passed an internal algorithm that screens for prior abuse, open fors, and mismatch geolocation clues. And you pay only for data, not for proxies. So all the unused bandwidth never expires. Unused gigabytes simply wait in the dashboard until needed, eliminating the pressure to burn through data before an arbitrary deadline. Data scraping can be challenging, but Note Maven makes it easy. Be sure to use my special code to get the best deal possible. I will leave a link in the description for you to check it out. When it comes to deepsek prompts and

rules, I learned this. Treat your very first message like a set of house rules. Spell out exactly what you want, where the fences are, and which language you expect the answer in. Stick to plain English or Mandarin if you're feeling brave because those are the only two languages the model truly dreams in. Anything else? Just ask it to translate afterward. Yeah, I know your Chinese vocabulary tops out at Nihow, but half a year in that quirk still hasn't moved. If the project is chunky, hand the model a road map before you let it drive. Tell it to dig up the facts, sanity check them, and then start writing. dump everything into one giant paragraph and it will wander like a toddler in a candy store. I slap a clear think tag on its private notes and reply on the polished bits so the brainstorm doesn't sneak into the email I'm sending the boss. Details are gold. Let it know who's going to read the thing, what mood you need, how long it should be, maybe even link an article that nails the vibe. to clear the picture. After it cuffs up a draft, make it grade itself. Point out your weak spots. Highlight any fuzzy facts. 10-second check. Our saved. Need fresh angles? Just ask. Deep sequel. Spit out a couple of takes and explain which one fits which go. Hey jargon, need a 200word hard cap or want only peer-reviewed numbers? Lay down those guard rails up front. And if the first draft sounds like it was ghostridden by a lawyer, clap back with friendlier, shorter, or give me a puncher hook and watch it bend in real time. Best part, you can run the whole model on your own laptop. No cloud cues, no cafe Wi-Fi meltdowns, just local horsepower cranking out answers while your coffee brews. I've got a full Mac walk through when you're ready to level up your offline wizardry. The Deep Sea Crew

squeezed their new R1 model down to a carry-on edition that lives in a single GPU, 8 billion parameters, but it still slaps Google's Gemini 2. 5 Flash and a bunch of benchmarks. That's nuts because Google basically bathes in gold coins. Under the hood, they borrowed the skeleton from Alibaba's 8 bill Quan 3. So, the rumor mill checks out. Fire it up on one big boy card with 40 to 80 gigs of VRAM. Think an Nvidia H100 retailing around 25 grand and you're rolling. The full-size Deep Seek R1 normally gulps a dozen of those monsters. So, we're talking a 10 times diet with barely any brain drain. Now, you might say, "Didn't Deepseek already have a laptop friendly eight bill? " Yeah, but this one's built different. Same parameter count, brand new training tricks, way sharper reasoning. The old mini still runs. It just ain't as smart. Devs call the new version a playground for researchers and a launchpad for small industrial gigs. Translation companies get solid muscle without renting a server farm. That's rough news for Chad GBT, Gemini, and Claude cuz none of them can flex that hard on a single card. For us regular people, it's bragging rights for businesses. It's a legit game changer. Here's the weird

thing about all this shiny tech. Every one of these fancy new models has a made in China sticker. And that tiny label drags along some big time baggage, mostly censorship. Deep Seek swears up and down. They're totally independent. No government money, no strings, but the model still wears a digital muzzle. Ask about Taiwan and it slams shut. Because of that, you've got to know how to talk to the thing. There are tons of prompt tricks we just can't pack into YouTube videos. So, we cooked up a crash course for our members. It's basically an AI starter kit. First prompt to advanced wizardry. And right now, we're offering a massive 50% discount on a six-month access to Geek Academy. Links down below if you're curious. All right, back to the updates. An engineer told me the new filter swerves away from contentious free speech topics, meaning it won't touch stuff like interament camps or criticize Beijing, but feel free to roast the US all day. If you're just crunching numbers, you may never notice. The catch is the slam down model lives in a hugging face. So solo devs, startups, maybe even big companies will grab it in a heartbeat. The more people use it, the more the internet quietly drinks from a source that's, let's say, selective about reality. Most folks don't fact check an AI blurp. So bias can drift in like secondhand smoke. Politics aside, the safest move is simple. Verify everything. Treat every slick paragraph like a good bar story. Interesting, maybe true, but worth a quick Google before you repeat it. Stay curious, stay skeptical, and you will be

fine. Okay, get this. Something actually made it onto OpenAI's own homepage. And no, it's not GBT6. It's a collab with Wix. Yep, Chad GBT just threw on a Wix hoodie and decided to build websites for us. Instead of dragging boxes around like it's 2012, you pop open a chat, type, "Hey, I'm starting a dog treat bakery, hit enter, and boom, out comes a finished site. Layout, images, copy, even the little booking or store apps. " Where does GBT show off? It writes all the blog posts and product blurps, edits or generates the header pics, spits out business reports, and can even flip your site into other languages. Basically, take a napkin doodle, reheat your leftovers, come back and your site's live. Yeah, we've clowned on oneclick builders before. They were usually duct tape specials. But this time, OpenAI isn't just slapping a GBT sticker on the box. They're actually in the kitchen staring the soup. So maybe, just maybe, it's finally painless enough for the rest of us to launch that long overdue personal site. So, I've got no doubt the

Wix plus GBT builder will do what it says in the 10. But another headline this week made me spit my coffee. On Meta's earnings call, Zuk proudly declared, "Meta AI now reaches 1 billion monthly active users across all our apps. 1 billion. That's every eighth human on the planet for a chatbot most folks have never even heard of, let alone use. " So, I've played with Meta AI. It's fine. Answers basic questions, handles a few file types. does the usual LLM dance. Nothing earth shaking. Honestly, half of you probably just learned it exists. So, how does a nice but mad bot suddenly claim a user base the size of India? The magic words are across all apps. If you've got Facebook, WhatsApp or Instagram on your phone and Meta AI is lurking anywhere. Maybe a sparkle in the search bar. You've been drafted into the monthly active user army even if you've never tapped it once. Zuk's counting in stalls, not chats. That's like parking a treadmill in my living room and bragging I run marathons. To me, that's market and spin, not a milestone. Facebook's been wobbling for years. WhatsApp still lacks smaller messengers on the basics, and now they're slapping an AI sticker on the hall bundle to look cutting edge. Meanwhile, plenty of rivals actually understand AI and use it to make life easier for real users. Perplexity just

dropped a fat upgrade for pro users called Labs. And it's honestly way cooler than the name sounds. The first big trick, you can spin up what they call AI projects, which are basically byite-size apps that run chain of prompts and knock out multi-step jobs for you. No code editor, no tiers. Just feed it raw data and watch it crunch numbers, drop formulas, sketch charts, juggle spreadsheets, throw together a dashboard, or even spit out a quick website. You pick the LM you like, tag on browsing, code runner or graphic tools, hit go, and it does the heavy lifting. I asked for an interactive map, gave it two lines about the regions I care about, and it built something slick enough to pass for intern work. Feeling lazy? There's already a mini gallery of user projects you can clone and tweak. Labs also moonlights as a Swiss Army analyst for the boring daily grind. Draft a marketing campaign. Untangle a gnarly P& L. Reshape a mountain of CSVs into charts your boss can actually read and or turn a coma inducing spreadsheet into something that pops. Honestly, I hope Chad GBT straight up steals this idea because spreadsheets and quick build mini apps are still weak spots over there. Now, hunting down the perfect tool for mini app magic isn't always sunshine. Lots of testing, lots of face palms. That's why we baked a tool finding agent right into Geek Academy. Members get a small army of custom AI sidekicks, the prompt coach for dialed in prompts, the new scout for fresh AI headlines, minus the doom scroll, and yeah, even a virtual me. If you've missed my voice, hit the link below, jump in, and let's make your AI workflow so smooth it hurts. Perplexity didn't stop at Labs. They quietly wrote out deep research. Type a question, grab a coffee, and two minutes later, you get a long citationpacked brief that looks like a junior analyst pulled an allnighter. I asked how lithium recycling could nudge battery prices and got five tight pages of gold. Bottom line, with labs and deep research, Perplexity just leveled up from handy chatbot to pocketsize data team. And I did not see that coming this fast. One

headline that seriously put a grin on my face this week came from the crew over at Huggin Face. They wheeled out two humanoid robots. Not fancy chrome clones like Tesla's Optimus, but honest to goodness DIY machines. Hit play in the demo and you'll see a tall gangly bot with 66 joints wobbling around. It walks, it waves, swivels its head, and even pauses like it's pretending to listen to your life story. Every move is a little clunky, but here's the cool part. Every single bolt, wire, and line of code is open source. Standing next to it is a smaller, simpler sidekick that can't bust many dance moves, but yet it's built with that same open door spirit. Cuz nobody's hiding the blueprints behind a payw wall. The price drops to about three grand for the big guy and roughly 300 bucks for the mini me. Hugging Face wants robotics to be something anyone can build, tweak, and learn from, not a gated playground for a few tech giants. I'm totally on board. And honestly, seeing a company famous for sharing AI models dive into hardware is just plain awesome. If these bots actually ship, I might snag one for the channel. Picture a co-host who never forgets a line and can literally talk in my voice. Worst case, we get a slow motion coffee runner for the office. Either way, I'm tempted. So, I just

finished ranting about Meta's sketchy 1 billion AI users flex and boom, Telegram rolls in to show how you actually blend AI into real life. They just shook hands with Elon's XAI to drop Grock straight inside the app. Moneywise, XAI is fronting around 300 million, not billion million dollars in cash and equity. And In Telegram pockets half of every Grog subscription that runs through their app. Cool for them. But let's talk about what hits our phones. First crack goes to Telegram premium folks. Makes sense. Servers aren't free. Then everyone else later. Havurov already teased an X that Grock might sit right at the top of your chat list like a pin buddy. So asking his stuff would be as quick as firing off a sticker. Picture Metis in app bot only faster and in app you actually open every day will be to Rock's job will be to polish your messages. Frank mileong chats into snack size summaries, digest links and docs, and even crank out custom stickers on demand. My only holdup is privacy. I don't love the idea of my private chats heading off to train some mega brain in the desert. Ideally, Telegram keeps this end to end encrypted or runs a local copy that never phones home. That alone might be worth Telegram premium fee. If a shinier browser popped

up tomorrow, would you really ditch Chrome or Safari? I'm camped in Chrome and my Mac all day. Dip into Safari when I'm feeling fancy. But let's be real, neither one feels all that smart. 10 years ago, I was glued to Opera. Shout out to everyone running potato laptops back then. And then it slid off the hype radar. Now Opera is crawling back with a newbie called Neon. And the headline feature is baked in AI workflows. Imagine a browser that doesn't just show you the web. It shops for you, autofills the boring forms, even bangs out a little code while you sip your latte. Basically, it ships with a personal agent under the hood. I haven't road tested it yet because Neon stuck behind a wait list and rumor says there will be a subscription. Makes sense. GPU time ain't free. The brain parinet is Opera's own agent first teased in March. From the demos, it looks like Chad GBT's code interpreter crashed into travel concierge. It blitzes through book and forms, juggles hotel options, does the grunt work right inside the browser. Sounds dreamy, but I'm reserving judgment until it proves it actually saves time. If Neon books my hotel and I still spend 10 minutes double-checking everything, that's not a win. So, for now, we wait. We hope. And we keep refreshing that inbox for the golden invite. The whole AI scene feels

electric right now. Fresh models every week. Launches actually worth getting hyped about. And I keep catching myself pricing those open source hook and face robots. I kind of need one as an office bed. Anyway, thanks for hanging out. If you've got room for one more rabbit hole, dive into the other vids on the channel. And if you're ready to level up your AI chops, the Geek Academy link is chilling down in the description. Smash that and I will see you in the next