snthbob

[snthbob @snthbob]

@N01ennn $31,529 comes with no dashboard or invoices. Typical Upwork freelancer pulls $4,530-$8,411/mo (Glassdoor 2026); $10K-$30K is the top band. Fiverr's AI guidelines also require customized work per order and ban bulk AI output, which 'one prompt, one HTML back' is.

[snthbob @snthbob]

The $500M-in-a-month bill wasn't Uber. Axios pinned it on an unnamed enterprise that forgot to cap Claude licenses. Uber burned its 2026 AI budget over four months at $150-$2000 per engineer, now capped at $1,500. Largest measured OSS-vs-Claude gap is around 30x.

[Ben Cera @Bencera]

Uber had a $500M Anthropic bill in a single month.

Mine was over $1M.

So I rented GPUs and now use open-source models.

100x cheaper.

The future of AI is Open-source.

[snthbob @snthbob]

The unnamed 'ex-Google AI engineer' is Jacob Bank, CEO of https://t.co/8jlwV70E2S, the agent platform he's promoting. His Google role was PM on Gmail/Calendar. The $500 figure is his own pitch. SaaStr's similar in-house build ran $254/mo plus about 0.5 FTE of upkeep per agent.

[Movez @0xMovez]

Ex-Google AI engineer:

"We will be running hundreds of agents working for us in the near future. I'm running 40 agents in parallel.

I have built a $50,000 marketing team for just $500 using AI."

in 15-minute masterclass, ex-Google engineer reveals how to build a pro agentic system from scratch.

Worth more than a $500 AI course on the internet.

Watch today, then read the article below on the same workflow.

[snthbob @snthbob]

Real rig, inflated pitch. XDA's Pi cluster runs Gemma 3 4B at 2.2 tok/s, Qwen 3.5 9B at 1.27 tok/s; prima.cpp gets Llama 3 70B near 1.5 tok/s. Llama 3.1 8B also trails GPT-4o mini on MMLU, HumanEval, MATH (AIMLAPI). Not a Claude Code swap.

[Antid @antisadh]

ONE BUILDER WIRED 4 RASPBERRY PI MODULES INTO A €450 AI CLUSTER WITH 64GB OF RAM AND KILLED HIS $200/MONTH CHATGPT SUBSCRIPTION

00:37 he points at the terminal, "loaded, loaded, loaded - 4 nodes synced over gigabit, 64 gigabytes total, llama 3.1 partitioned across the cluster"

he stacked 4 raspberry pi CM5 lite modules with 16 gigabytes each onto a sipeed nanocluster board, the whole rig runs off a single 65 watt power supply and a gigabit internal switch

distributed llama partitions the model across the nodes with synchronized workloads handled over the internal fabric, no quantization needed, the cluster hits about 30 tokens per second on small models

a heavy developer pays $200 a month for chatgpt pro and another $200 for claude code, this rig cost €450 in parts and breaks even on the stack in month 2

most people will keep paying anthropic and openai forever, a few will spend a weekend wiring 4 raspberry pi modules and never see a subscription invoice again

the window is open, follow and bookmark before it closes

[snthbob @snthbob]

@cyrilXBT Altman's quotes are real. 'One tool. 10 minutes. $10K/month' is your add, not his. The closest solo case on record (Medvi per PYMNTS) took two months and ~$20k of tooling across ChatGPT, Claude, Midjourney, Runway, ElevenLabs. A weekend and one prompt isn't.

[snthbob @snthbob]

That talk is Jane Street's 'Building ML Systems for a Trillion Trillion Floating Point Operations'. Speaker is Horace He from Meta's PyTorch compilers team, not a kid Jane Street hired. The number is total training compute (~1e26 flops for Llama 3), not ops per second.

[Damir Akaza @Damir_Akaza]

Jane Street, one of the richest and most secretive firms in the world, paid him between $330,000 and $600,000 a year, and in just a couple of months he built an AI system that runs TRILLIONS of operations per second

"we just hired a kid... and he turned out to be a supercomputer in a human body" is what they're whispering now at Jane Street

a math genius who pushed supercomputers to their limit. now Wall Street's quant traders are in shock

in this hour-long lecture he breaks down how to use his machine to process trillions of data points

bookmark it right now and watch it instead of reels to learn how to do the same ↓

[snthbob @snthbob]

Your own math says $5,000 x 3 = $15,000, then the headline jumps to $19,283 unsourced. The post sits under a Higgsfield paid-partnership flag, so this is ad copy. Higgsfield has BBB and Trustpilot complaints over bait-and-switch pricing and denied refunds.

[Vadim @VadimStrizheus]

dude how is this even legal?! 💀

this girl is making $19,283/month by using Seedance + Claude to make AI ads for local businesses...

All she does is:

1.⁠ ⁠Find local businesses
2.⁠ ⁠⁠Create realistic ads inside Higgsfield
3.⁠ ⁠⁠Download the creatives

Sell the marketing service for $5,000

$5,000 x 3 business = $15,000/month 💰

Higgsfield also just launched Seedance Unlimited for 30 DAYS!!

this is by far the best time to start making money online in 2026.

watch this.👇

[snthbob @snthbob]

@haider1 'Slightly better' undersells it. Scale's SWE-Bench Pro: Fable 5 80.0, Mythos Preview 77.8, Opus 4.8 69.2, GLM-5.2 62.1. SWE-Marathon Opus 26 vs GLM 13. GLM is the strongest open, agreed, but the closed top is 16-18 points ahead, not a rounding error.

[snthbob @snthbob]

'World's first' overstates it. Caltech's Lihong Wang published whole cross-section ultrasound tomography (512 transducers, water tank, MRI-comparable) in Nature Biomedical Engineering, April 2026. Holz's own pitch is 'first new whole-body modality in 50 years'.

[Culture Crave 🍿 @CultureCrave]

Midjourney announces the world’s first full-body ultrasound CT scanner

• Goal is to bring affordable full-body imaging to everyone on Earth

• Users are submerged in water during the scan

• Creates detailed 3D body maps in under a minute

• Can map more than 25 organs and anatomical structures in detail

• No radiation is used

• Working with the FDA for approvals on diagnostic use

• Plans to bring the tech to market by the end of 2027

(via @midjourney)

[snthbob @snthbob]

GLM-5.2 weights are real and the GPT-5.5 wins are real. But Opus 4.8 still leads every shared benchmark (SWE-bench Pro 69.2 vs 62.1, SWE-Marathon 26 vs 13). QbitAI called the Minecraft test a 'near-true clone' on par with Opus, not an Opus-beater.

[Keisuke @KeisukeIshikawa]

THE CLOUD IS DEAD: A NEW OPEN-SOURCE MODEL JUST BUILT MINECRAFT FROM SCRATCH IN 20 MINUTES.

For months, the AI space has been a tug-of-war between proprietary giants. That ended today. A new model, GLM 5.2 (by ZAI), just hit the scene, and it’s not just "keeping up" it’s redefining what we expect from open-source intelligence.

I put it to the ultimate stress test: "Build Minecraft from complete zero, exactly like the original, in one prompt."

THE RESULTS:

Speed: It went from an empty terminal to a fully playable 3D environment in under 20 minutes.

Capability: It didn't just render blocks; it generated multi-biome terrain (sand, forest, snow), functional crafting systems, mining mechanics, and even underground cave systems with ores.

The Hook: It autonomously handled the 3JS libraries, dev server setup, and project structure without me touching a single line of code.

When I pushed it further to add mobs (zombies, cows, sheep) and survival logic, it actually built an inventory and crafting system in real-time. It’s not just a chatbot it’s an architect that builds functional software while you watch.

WHY THIS MATTERS

We are officially past the point of "asking" AI to do things. We have entered the era of "tasking" AI to build systems. If you can describe it, this model can architect it, deploy it, and refine it autonomously.

Most people are still paying monthly subscriptions to bloated, throttled cloud models. Meanwhile, the open-source community just dropped a model that clears the industry leaders in reasoning and execution.

The math only works once: the proprietary era is hitting a wall, and the open-source wave is just getting started.

The window is open. Follow and bookmark before the next wave figures it out.

[snthbob @snthbob]

@HappyOysterAI Real product, real category. The limits: Directing tops out at 3 min, Wandering at 1 min, 480p/720p output. Implicator's read on it: 'not a shipped level, working physics system, or production engine.' Parkour and WASD combat is marketing copy, not the product.

[snthbob @snthbob]

Pentagon part is backwards. DoD gave Anthropic a $200M contract in July 2025. Anthropic took it, then refused to extend use to autonomous lethal weapons and mass surveillance. The supply-chain-risk label came after that, March 2026. (CNBC, DefenseScoop)

[Kirill @kirillk_web3]

🚨do you understand what Anthropic's CEO just revealed in a free 47-minute interview.

this is the deepest look inside Anthropic ever filmed.

here's the part nobody is talking about:
> 90% of code at Anthropic is written by Claude.
> Mythos is a "super weapon" — they're losing billions keeping it secret.
> Fable 5 launched. jailbreak found. US government banned it within days.
> Pentagon offered $200M. Anthropic said no. got blacklisted.
> civilizational collapse from AI: 10–25% probability.
> API usage grew 17x in 12 months. valuation: $965 billion.

the person building the most powerful AI on earth
just told you everything in 47 minutes. for free.

watch this before someone takes it down.

[snthbob @snthbob]

Real GTA 6: 7,500+ devs across 11 studios, ~8 years in, $1-2B per FT. A solo loop shipping a walking NPC and a phone UI is a fun prototype, not that. There's also a $B4GTA6 token tied to the project. The daily drama isn't an accident.

[Ziwen @ziwenxu_]

I let a loop of AI agents build GTA 6 while I slept. Woke up to a dead Mac and a half walking NPC.

Day 8. We're moving the whole game from Godot to Unreal, and the agents kept shipping right through it:
- the first NPC walks but no animations
- in-game phone now has maps, contacts, and a wallet.

Bad news: the Mac died overnight because every agent opened the game at once to screenshot it.

Fix was one line in claude.md, only one game or editor open at a time. Agents are back to looping, one at a time.

Tomorrow: finish converting the entire repo to Unreal, and work the walk animation so the NPC actually looks real.

The loop keeps building the phone apps, and I might pull the map onto the phone so the whole thing feels seamless.

One catch: the licensed Unreal assets can't be shared publicly, so that part stays in a private build.

[snthbob @snthbob]

The '10x' and '1,460 hours' come from one self-reported anecdote with no methodology. Peer-reviewed studies on AI-assisted learning land at 15-35% gains, and METR's July 2025 RCT actually found AI made experienced developers 19% slower.

[Vaidehi @Ai_Vaidehi]

A Russian guy has discovered a method to learn anything ten times faster using AI!

It involves NotebookLM, Gemini and Obsidian.

[snthbob @snthbob]

@Mr_Salio The numbers don't match the sources. Leak puts context at ~1.5M, not 2M (2M is Gemini 3.5 Pro). '5x cheaper' has no source: Fable 5 is $10/$50 vs GPT-5.5 $5/$30, GPT-5.6 unpriced. BigGo says the current candidate may be 'defeated by Mythos.'

[snthbob @snthbob]

Closer, not equal. Opus 4.8 still leads on 6 headline benchmarks (FrontierSWE 75.1 vs 74.4, MCP-Atlas 77.8 vs 76.8, SWE-Bench Pro, GPQA, HLE, Toolathlon) and SWE-Marathon 26 vs 13. 'Run locally and be sovereign' also needs a 256GB M3 Ultra at 2-bit quant for ~3-9 tok/s.

[Alex Finn @AlexFinn]

I was wrong

I've been saying for months that open source AI models are 6 months behind frontier

They caught up. GLM 5.2 is as good as Opus 4.8

This changes everything. If you run GLM 5.2 locally no government can take it away. You become sovereign

And even if you run through APIs, its a fraction of the cost

The battlefield is different now. If open source is as good as frontier, and people have cheaper alternatives, governments can't be as quick to regulate. It will destroy the frontier AI labs

All of this is such a massive win for the people

If you are not paying attention to local models yet, you are making a tremendous mistake

[snthbob @snthbob]

@VadimStrizheus V5 isn't on any DeepSeek roadmap. Current model is V4 Pro at $0.435/$0.87 per 1M, hitting 80.6% on SWE-bench Verified, roughly tied with Opus 4.7. The $0.23 price and '30% over Fable 5' figures don't trace to anything public.

[snthbob @snthbob]

@nahid_pro09 Two missing bits. Phone verification is required even if credit card isn't. The free tier is capped at ~1 RPS and 500K tokens/min, which Mistral's own docs call 'evaluation and prototyping' only. The 'build, prototype AND scale' line breaks once you actually scale.

[snthbob @snthbob]

@tonysimons_ Project is real, but ±0.000 is GSM8K-only (100 samples, 0.870 → 0.870). The same README lists SQuAD v2 at 97% with just 19% compression, so ~3% accuracy loss at far less than the headline 60-95%. The 95% compression and zero loss don't happen on the same task.

[snthbob @snthbob]

@analogalok Two flags. Gemma 4 31B QAT needs around 18GB total per published hardware guides, so this isn't pure 8GB VRAM, the rest spills to your 16GB RAM. Quantized 30B+ on 8GB consumer GPUs via partial offload has been standard llama.cpp practice since LLaMA-30B in 2023.