Spencer

Dunno but the way things are going sounds like we're looking at better and better efficiency over time Gemma 4 26B is a pretty competent model given that it can run on a medium-high end laptop. Have a friend running DeepSeek with Claude Code doing pretty demanding things (and getting great results) and is spending under $5/day on API calls Sonnet 4.6 can generally perform similar to Opus 4.6, allegedly with significantly less compute I'd imagine they'll mostly smooth it out before subsidization runs dry

Exactly how subsidized are tokens for Claude, Perplexity etc right now? Realistically if we're building workflows dependent on these that are running two $200/m max plans... does this change to... $1000? $5000? More? When?

Was fed up with ad-ridden YouTube to mp3 conversion websites so I created my own in 10 minutes. 10x better experience than everything out there. Faster too. Made it uniquely to my preference. Don't forget that you can literally just vibecode simple utility apps for yourself.

It cost me fking $6.61 just to set up a design system in Claude Design No, not to create anything. Just to add brand assets and let it process (Oh, and this is the price before any revisions) Burned my week limit in literally 2 hours ($100/mo plan). Now onto extra usage. If this is the VC-subsidized "golden era" of AI and it still costs this much, everyone who gets too reliant on these tools are about to be fucked LOL

@KamStaszewski i mean its just an open source project that got overhyped not really a scam lol

@rendererr TIL: openclaw runs like any mlm scam on naive people, like our older relatives

Feels like every few years something gets yanked out of niche tech circles and anointed as the current speculative mania. I remember scrolling on here when OpenClaw started gaining serious traction and nobody could name a single truly interesting use case for it. Meanwhile I watched as my local MicroCenter sold 45 Mac Minis in a day. I seriously almost bought one out of pure fomo. Then when I was in China last month, there were free events to help grandmas get their OpenClaws set up. That's when I knew 100% it was a fad. "But it texts you first!" Why the fk do you want your AI to text you? Put the mac mini in the bag bro

@lukaz3v at the time i think gemini 3 was pretty solid (it's since taken a nosedive and is hardly usable) idk if were talking about the same thing but im referring to the types of tech fads where your parents start asking you about it because they saw it on facebook

@rendererr humans get caught up in hype a lot, did you see the gemini 3 stuff at the end of last year lol I don't think that's such a bad thing though (in general)

@johnwalters_sd Yeah there's some neat tinkering/hobbyist use cases but it's nowhere near groundbreaking as the hype would suggest

@rendererr Saw some guy's post bragging how he was able to just talk to it while he was on the phone driving his car. He was telling it how to configure some AI model and or where his keys were or something. But still not saying what he'd ever actually accomplished with it.

I've been messing with Perplexity lately and it honestly beats Opus 4.6 in a few very specific ways I wasn't expecting: When you're trying to gather a general sentiment on something recent, I find Claude can't reliably access brand new information. Perplexity almost feels like a superpowered version of the Google AI Overview feature that everyone hates. Personally I found it impressive at helping me come up with an investment strategy, because it took current events into account up to the past hour. Claude feels more powerful in general, but its ability to gather current info is hit-or-miss and I've found it can sometimes leave out important information when researching. If you already have Claude, Perplexity is definitely also worth giving a try.

Coding with AI should matter to you even if you can't code. I got 8m views in 33 days on a fresh Instagram account after only 15 posts thanks to Claude Code. This was my strategy: •I started by saving content that's already performing well in my niche. In this example, my niche was cars, specifically the Mazda Miata. •Once I had some initial high-performing content examples saved, I asked Claude Code to deep-research common tropes that appear in the car/Miata niche until I had a full list: "slow", "small", "feminine", "absurd modifications", "neighbors hating noisy cars", etc. •I used all this data with Claude Code to create a Tinder-like app to help me generate viable content ideas. I can easily left-swipe to reject, right-swipe to approve, and optionally leave comments describing why Claude's ideas are good or bad. (I know effectively nothing about coding, by the way) •Each swipe improves the quality of the ideas my app generates by essentially batch-saving my responses as a json file and feeding it back into the idea generator until I'm left with a list of strong content ideas to produce. This allows me to stay on track and focus exclusively on creation. The lesson here is that the ability to distinguish good and bad content is one of the only remaining moats now; AI fundamentally changes the creative process by giving you abundance first. Your distinct advantage as a human is to have taste. Everyone just wants to talk about automation (which can be great, no doubt), but don't forget to build systems like these for whatever you're doing. The mindset around creating an 'app' needs a fundamental change. Releasing or directly monetizing the app itself is not always the main priority anymore. Building highly specific apps to use internally for your business or personal life is the future, and will continue growing more and more accessible as tools get better.

Another horrible day to be Adobe, Figma, Lovable, and anyone who isn't Anthropic themselves Polymarket says 90% chance we're getting Claude Opus 4.7 by the end of this month (or even possibly this week), and design stocks are already reacting Leaks suggest Anthropic is testing an alternative to Claude Code that would directly rival Lovable (who, by the way, is currently powered by Claude Opus 4.6 on their backend) This just makes it easier and easier for normal people to build websites and web apps without any coding skills, and without facing intimidating interfaces Truly never a dull moment

Got like 6m views in my first month posting on instagram reels as a fun side project This niche is sooo underserved given how many people watch car reels. Massive supply:demand discrepancy Feels like everyone making car reels is a midwest middle schooler who just got their first camera. No attention to detail or originality - often horrible taste too. Incredibly easy to stand out if you know anything about short form content and are calibrated around the culture. I drive a completely stock 2016 Miata ($20k car) and had video after video go astronomical

Claude will make you seriously RETARDED if you're not careful Just vibecoded a scheduling app for a local service business client to avoid paying monthly for an equivalent SaaS Fucker one shotted it first try Got to the part where I have to "configure OAuth in Google Cloud" or something, whatever that means Instead of figuring it out myself I just gave Claude full access to my computer so it could click around and do it all for me while I put away the dishes 10/10 didn't need to use my brain once. Not a single moment of discomfort Good times create weak men

Claude just did all my taxes while I got drunk at brunch (tax deadline is in 2 days btw) Step 1: Download PDFs for all sources of income and expenses (Stripe 1099s, Robinhood investments, business credit card etc.) Move everything to one folder Step 2: Download Claude desktop and load up Claude Cowork; connect it to the folder with your PDFs. Use a compatible browser to log in to your platform of choice e.g., FreeTaxUSA Step 3: Tell Claude to use your computer and your forms to complete all your taxes, but tell it NOT to submit (you'll want to give final review). Cowork will be able to control your screen; click around, and type Step 3.5: Also tell it to compile a list of questions for you while it's working to help maximize writeoffs. It found me over $2k in savings that I would have otherwise missed Step 4: Go enjoy your day while the robot handles the grunt work for you. Welcome to the singularity

If you don't care about anything try --dangerously-skip-permissions mode in Claude Code Pros: No more spamming enter Cons: Might format your hard drive and send your data to Palantir in a .zip file

swallowed the Zyn

Moltbook really made a "Social Network for AI Agents" as if we didn't already have Linkedin

This truck animation would have cost THOUSANDS and now anyone can make it for pennies. The interactive map would have required diverse coding knowledge; tediously mapping coordinates to draw the perimeter and putting the coordinates into code. Now anyone can make it for pennies. The "Process" animation would have required advanced Figma and CSS skills. Now anyone can make it for pennies. Claude Code completely removes the technical limitations of adding advanced components to websites. This is available now.

Important reality check on the local AI hype since everyone says we have Opus 4.6 at home now: Spoiler - the caveats around local AI make it hard to justify beyond a hobbyist's use case. I say this as the owner of an M5 MacBook Pro w/ 48GB RAM and a PC with an RTX 5090; $3,100 and $5,000 respectively. Gemma 4 hosted on LMstudio with Claude Code generated an objectively solid multi-page website (without internet connection) for a mockup coffee shop called Gemma Coffee. Gemma used the Claude Code tools correctly and there were minimal hiccups, but my main epiphany is that the $3,100 MacBook is simply nowhere near enough to get any serious work done. For this project I had to sacrifice some intelligence and opt for the 26B model size over the 31B. Even then, I was having to eject the model from LMstudio and reset Claude Code to free up memory after 1 or 2 prompts. It would generate the initial code with little trouble, but making edits requires dumping all of the code back into the context window. You can see why this would eat up the little wiggle room I had to begin with. I was also doing all of this with a conservative 85k token limit (under 1/10 of what Opus currently has), and the RAM was still just filling up too fast.. Just because you can run a model in Ollama and fire off a few chats in your terminal doesn't mean you can connect the same model to Claude Code and get work done. Until the overlords can pull some strings and get this running with less hardware I'll just be over here sending $100 per month to Anthropic. (Here's the site I made for this project btw)

Upgrading from a 14" to a 16" MacBook

If you want a new MacBook, get it now. This is your last opportunity to get this much RAM at current prices. Apple secures their price agreements way in advance, and their current RAM deal expired in January; right before the new M5 Pro/Max models dropped. With a redesign on the way later this year, they're gonna take this opportunity to direct your attention to something new and shiny while quietly hiking up RAM prices to match the market. Just grabbed the 16 inch M5 Pro with 48GB RAM at Microcenter for $2789 ($310 discount). Recent open source models like Gemma 4 and Qwen 3.6 have me bullish on local AI in the coming years. This won't last forever. Secure as much RAM as you can. You're gonna want it.