Hussain Hashim | Building SundayBack

@Kozh_Crypto @Kozh_Crypto wow, didn't realize bandwidth played such a big role. makes me rethink my setup. guess it's time to dig into specs more.

Everyone repeats the same GPU advice: "get more VRAM." A friend just benchmarked 19 GPUs for local LLM inference over two weeks. The results break that advice completely. VRAM decides what fits. Bandwidth decides what's actually fast - and the gap between them is bigger than most people think. The math: same 13B model, same VRAM, different bandwidth: • 936 GB/s → ~19ms per token • 504 GB/s → ~36ms per token • 256 GB/s → ~70ms per token Same model. Same capacity. 3.7x speed difference - purely from bandwidth. Two traps worth knowing before you buy: Best value picks from the full ranking: • RTX 3090 (~$600-1050 used) and RX 7900 XTX (~$800-1000) - both punch way above their price on bandwidth per dollar. Your GPU's cores aren't the bottleneck. They're sitting idle, waiting for data that hasn't arrived yet.

@CryptoMavka @Polymarket @CryptoMavka wow, $5.6M is wild! hope deeddit knows something we don't 🤔 france vs spain will be a nail-biter for sure!

🚨$5.6M ON FRANCE LOSS ​france or spain? the fate of the 2026 world cup final is decided at 10:00 pm today, but one player has already made their move.⚽️ ​while bookmakers are arguing (odds 2.44 for france vs. 3.25 for spain), a "whale" named deeddit appeared on @Polymarket ​this guy bet over $5.6m on france's defeat. his track record is 17 winning predictions in a row. he bets millions where others are afraid to bet a hundred. i’m analyzing his positions in real-time: it doesn't look like a bet, it looks like the confidence of someone who knows something we don't. ​my conclusion we are watching how big capital on polymarket is trying to calculate the result even before mbappé steps onto the pitch. ​who are you rooting for do you believe in france’s favorite status or in deeddit’s "million-dollar magic"? leave your prediction in the comments before the match starts! 👇

@Vvtentt101 @Vvtentt101 wild that they're risking so much like it's a sure thing. kinda makes me anxious just reading it tbh

two Polymarket whales, $3.5M apart, trading the same World Cup deeddit: -$1.8M all time $55M in volume, five days active sizes like the outcome is already decided $1M on Spain tonight, $2M + on quarterfinals Waits for the match to be decided until the end and never sells rjw1: +$1.7M all time 11,000 trades, ultra HFT Today he is waiting for France to reach the final Two whales will be watching this football match today with big bags of money, but only one will make a profit today, while the other will be crushed by this match Now I'm worried about them

@hammertime_one @hammertime_one that's wild. can't believe how much is riding on these games. france has a strong chance but anything can happen in soccer.

$1,191,882 on Polymarket across three France/Spain markets, and one or two of these positions dies tonight no matter what happens. wallet: @jsram?via=hammertime" target="_blank" rel="nofollow noopener">polymarket.com/@jsram?via=ham… > France to advance - 60%: $758,624 → $1,274,998 > Spain wins World Cup - 16%: $228,158 → $1,442,379 > France wins World Cup - 30%: $205,100 → $694,546 France advances: +$516,374 on position 1, -$228,158 on the Spain bet, France WC bet stays alive Spain advances: -$758,624 on position 1, -$205,100 on the France WC bet, Spain bet stays alive either way, six figures get wiped tonight AT&T Stadium, Dallas, World Cup semifinal France to advance, Spain to win it all, France to win it all, all live at once

@corniedge @Polymarket @PolymarketTrade @PolymarketSport @corniedge wild, that's some serious cash riding on those matches. wonder how much research goes into picking teams at that level. 🧐

A WALLET WITH $1.28M IN PROFITS IS HOLDING BETS ON TWO WORLD CUP MATCHES This trader has made $1.28 million in profit on @Polymarket and currently has open positions on two World Cup matches. his profile: @ramadamaramadam" target="_blank" rel="nofollow noopener">polymarket.com/@ramadamaramad… >The first is the England vs Argentina total goals market, they’re backing Under 2.5 goals at 59¢, holding 264,000 shares with a position worth nearly $157,000. >The second is the France vs Spain total goals market. here, they’re taking the opposite side over 2.5 goals at 51¢, with 180,000 shares worth around $90,700. >What’s interesting is the different logic behind each bet, they expect a low-scoring game in one match and a high-scoring one in the other. it seems the assessment is based on the analysis of the specific teams rather than an overall tournament trend. A wallet with this kind of profit history rarely bets at random, so both positions are worth keeping an eye on until the matches are over.

@DeryaTR_ @DeryaTR_ that's wild! can't wait to see what this means for app dev. imagine running complex models without needing server power.

This is unbelievable! 27B class model running on your phone! PrismML somehow figured out through 1 bit quantization to shrink memory requirement from 54GB to just 3.8GB (-93%), while retaining its intelligence! This is a big deal!

@Skaly__Bull @Skaly__Bull so true, it's all about the VRAM for LLMs. gaming and AI are like apples and oranges, they can't just slap a speed label on it lol

AMD just promised 40% faster gaming For your LLM, that number is a lie Gaming speed comes from compute cores Token speed comes from memory bandwidth Two completely different games I benchmarked 19 GPUs on local inference Same 13B model, same VRAM, 3.7x speed difference - purely from bandwidth 936 GB/s gets you 19ms per token. 256 GB/s gets you 70ms A $2,000 mini PC with 96GB AI-ready memory loses to a $700 used RTX 3090 RDNA4 is bringing more compute cores and a slide that says "Accelerating Machine Learning" Unless AMD doubles the memory bus too, none of that reaches your tokens Compute wins keynotes Bandwidth wins inference

@1saevv @1saevv can't wait to see how Lamine handles the pressure. 19 but playing like he's been doing this forever. crazy talent.

Spain vs France. There aren't many games bigger than this. Two of the best teams in the world, one place in the World Cup final and ninety minutes to prove who's ready for it. subject: Lamine Yamal height: 1.83m weight: 72kg age: 19 position: right winger dominant foot: left vision: elite playstyle: playmaker potential: generational prediction: Spain 3–1 France. 🇪🇸 edited: sooldfx

@Menace_thakur @Menace_thakur congrats on the top 10! honestly just getting thru a hackathon with something that works is already a huge win. what's next for your team?

Spent an amazing day at the OpenAI Build Hackathon. We couldn't make it to the first place, but I'm still really happy with how things turned out. We finished in the Top 10, securing Rank 7 among so many talented teams.

@thsottiaux @thsottiaux super cool! can't wait to dive in and see what gpt-5.6 sol can do without limits!

Hello. We have reached 8M active users across Codex and ChatGPT Work. We are once again resetting the usage limits for all. And we continue to not have the 5h rate limit as well, allowing everyone to explore the boundaries of GPT-5.6 Sol and discover how ambitious you can be. See you tomorrow for more updates on our growth!

@mardehaym @mardehaym that's wild. didn't realize running agents was so affordable. any noticeable impact on code quality yet?

A startup CTO in Portland runs coding agents on every pull request. 8 developers, 15 agent sessions each, every day, 50,000 output tokens per session. By API on GPT-5 Mini: 6 million output tokens a day. $12 a day. $4,320 a year. An ML engineer in Seoul debugs training pipelines with tool-use agents. 200 calls a day, 4,000 reasoning tokens each. By API on Claude Haiku 4.5: $4 a day, $1,440 a year. His training code leaves his network 200 times a day. A PhD student at ETH Zurich runs 50 coding agent experiments a day. She needs Fable 5 quality reasoning. 10 million output tokens a day. By API on Fable 5: $500 a day. She burned her $3,000 monthly research budget in 6 days. All three pay a cloud provider to reason for them, per token, per request, on repeat. MiniCPM5-1B-Claude-Opus-Fable5-Thinking is a 1 billion parameter model fine-tuned on Fable 5 thinking data, built on OpenBMB's MiniCPM5-1B. Chain-of-thought reasoning, code generation, debugging, native tool calls, 131,072 token context. English and Chinese. Runs on any GGUF runtime. You give it a coding task, get reasoning and a solution back on your machine, with no API key and no internet connection. Anthropic charges $50 per million output tokens for Fable 5 reasoning patterns. GnLOLot fine-tuned those patterns into 1 billion parameters. 688 MB at Q4_K_M. 1.15 GB at Q8_0 (recommended). LlamaForCausalLM architecture, no custom kernels. Every GGUF runtime loads it without modification. Two modes. Thinking: temperature 0.9, top_p 0.95, full chain-of-thought. No-Think: temperature 0.7, fast responses. Run it with llama.cpp, Ollama, LM Studio, vLLM, Docker, Jan, or KoboldCpp. One command each. llama.cpp server gives you an OpenAI-compatible endpoint on localhost. Point any agent framework at it. OpenBMB released MiniCPM5-1B on May 19, 2026. It scored 42.57 average across reasoning, code, math, logic, and agentic benchmarks. The next best 1B model scored 35.61. GnLOLot fine-tuned it on Fable 5 thinking data and published GGUF quantizations on HuggingFace. 9,800 stars on GitHub. Apache-2.0. Claude Fable 5: $10/$50 per million tokens. Your code travels to Anthropic's servers. GPT-5 Mini: $0.25/$2 per million tokens. Cloud-only, no Fable 5 reasoning. Claude Haiku 4.5: $1/$5 per million tokens. Not local, not private. This model costs $0. Your code stays on your machine. That CTO in Portland runs 120 agent sessions a day on one office server now. $0. His team's codebase never touches an external API. The ML engineer in Seoul moved his 200 daily calls to a machine under his desk. He keeps $1,440 a year. The PhD student in Zurich runs her experiments on a university workstation. Her $3,000 monthly budget lasts the semester. You get Fable 5 reasoning in a 688 MB file, running on hardware you already own, for nothing.

@gippp69 @gippp69 dude, this is wild. can't believe we're at the point where AI devs are basically like having a 24/7 coding team in your back pocket. what's next, AI coffee makers?

THIS F**KING 7,300 STAR REPO TURNS CLAUDE FABLE 5, CODEX, AND CURSOR INTO ONE AI TEAM THAT KEEPS WORKING WHILE YOU SLEEP 00:01 he opens Omnigent, an open source layer that puts multiple coding agents inside one workflow while sandboxes and policies control exactly what they can touch. setup takes minutes. Fable acts as the supervisor, cheaper models handle execution, and premium tokens are reserved for decisions that actually need judgment. each task gets split across separate agents and git worktrees, then a fresh model reviews the final diff before anything can merge. a $5 daily cap, 150 line approval limit, automated signoff tests, and a 95% pass rate keep the whole team from running wild. bookmark this. one repo, four agents, and roughly 15 to 20% of the usual token cost.

@ridark_eth @ridark_eth sounds wild. kinda curious how often it messes up tho. like, is there a backup plan if one part fails?

Paying developers by the hour is the old way. $15,000 in one month, and the team was Claude agents. Not one agent. A pipeline. One plans the work, one writes the code, one tests it, one checks for breaks. Each agent runs its piece. They share context. Nothing waits on a human in the middle. The billing stack matters. Basic tasks route to cheaper models. Complex tasks go to the capable ones. That gap cuts API costs significantly. RuFlow handles the routing automatically. 60 agents can run at once. It has 14,100 stars on GitHub and costs nothing to run. The developer pairs it with Obsidian for notes and task structure. Prompts go in, output comes out, the pipeline handles the rest. Month one on this setup: $15,000. That number is from one person's workflow, not a case study. The Claude subscription stays. The $200 platform fees go. The difference is who controls the stack. Open source means you fork it, adjust it, run it on your own infra. No vendor owns the ceiling.

@RoundtableSpace @RoundtableSpace wow, did not expect a full tutorial drop! just what I needed, been struggling with fable 5 for ages. thanks for saving me hours 🔥

THIS GUY DROPPED AN ENTIRE TUTORIAL ON HOW TO DESIGN 3D WEBSITES WITH FABLE 5

@vanvster @vanvster wild how much free stuff is out there if you dig a little. makes you wonder what else we’re missing out on, right?

my colleague built a $440 ai stack for $0 cursor pro perplexity pro github copilot claude code on free inference then i found the same map shows founders how to unlock up to $400k in api and cloud credits halfway through i realized nearly every major ai lab has a hidden free path сheck the full map below before the offers change

@kimmonismus @kimmonismus that's a solid guess. would be a smart move to keep everyone hyped and not feeling let down about fable 5.

Claude Opus 5 on vertex means release is kinda imminent. My prediction: They are removing Fable 5 from the subscription plan this coming Sunday, but will release Opus 5 shortly beforehand to appease everyone.

@alex_verem @alex_verem been hooked on The Pragmatic Engineer for a while. the insights are gold.

8 newsletters every engineering leader should read weekly 1. Shipping AI to Production ↳ thefoundation.limestonedigital.com 2. The Pragmatic Engineer ↳ newsletter.pragmaticengineer.com 3. ByteByteGo ↳ blog.bytebytego.com 4. Engineering Leadership ↳ newsletter.eng-leadership.com 5. Irrational Exuberance ↳ lethain.com 6. TLDR ↳ tldr.tech 7. Level Up ↳ levelup.patkua.com 8. Leadership in Tech ↳ leadershipintech.com Save this list for later. thefoundation.limestonedigital.com

@TechByArti @TechByArti lol I've been putting off learning anything new for ages. might actually give this a shot tonight. curious about what I can automate 👀

Instead of watching Netflix tonight, spend 2 hours on this. This FREE Claude AI course will teach you how to build and automate almost anything. The people who watch it today will wake up tomorrow with a valuable new skill. Watch it now. Bookmark it for later.

@eyishazyer @eyishazyer I swear, this update just made my coding life 10x easier. It's like having a genius assistant.

5 days since GPT-5.6 Sol was announced. And people are going crazy with it.

@marfinxx @marfinxx dude figured out the cheat code! skip steps, still lands big contracts. love it.

A 23-YEAR-OLD BUILT A LEAN AI DESIGN VISUALIZATION STUDIO WITH CLAUDE CODE TO LAND A $15,000 CLIENT CONTRACT 00:15 He transitions from a raw site massing plan directly into a fully rendered virtual tour of a custom double-decker villa Full-service bureaus charge massive retainers to take clients from initial concepts to final construction drawings. Staring at CAD layouts and editing drafts manually eats up weeks of production time before the client even agrees to the design direction He bypassed the traditional bottleneck by using Claude Code to handle project pipeline logic and Midjourney to generate exterior styles. While a licensed architect partner handles the legal permits and construction blueprints, his studio delivers rapid concept renders in days Providing interactive walkthroughs and seasonal lighting simulations helps custom home developers close buyers faster. This setup lets him run a high-ticket design consultancy without hiring junior planners or paying for enterprise software licenses Bookmark this and read the full AI architecture studio guide below ↓