😲 GPT-5 Initial Reactions

OpenAI’s GPT-5 arrived on August 7th, and Sam Altman wasted no time dubbing it “the best model in the world.” The new flagship now auto-routes between a fast core for quick chats and a thinking variant for tougher puzzles, eliminating the old model picker clutter. Early testers say the router largely feels invisible, until you ask it to “think hard,” when answers pause, then return with richer reasoning.

🧠 What’s Actually New?

• Massive memory: GPT-5 tracks up to 256K tokens (roughly a 700-page novel) without losing the thread, a big jump from GPT-4o’s 12K.
  
  

• Benchmark bump: On the notoriously tricky SWE-Bench Verified coding test it scores 74.9%, edging past both o3 (69.1%) and Claude Opus 4.1 (74.5%) while using fewer tool calls.
  
  

• Built-in personas: Four opt-in “personalities”—Cynic, Robot, Listener, Nerd—let users nudge tone without prompt-engineering.
  
  

• Pricing shuffle: Devs now pay $1.25/M input tokens and $10/M output tokens—cheaper than GPT-4o for inputs, pricier for outputs, and well below Claude’s top tier.

📉 Less Hallucinations but They Still Exist

OpenAI’s system card shows concrete gains:

• GPT-5-main produces 26% fewer hallucinated claims than GPT-4o.

• GPT-5-thinking cuts hallucinations 65% vs o3 and slashes responses with any major factual error by 78%.
  

• On anonymized production traffic, answers flagged as factually wrong drop about 45% compared to GPT-4o and nearly 80% versus o3 when the thinking mode kicks in.

That still isn’t zero, especially in low-resource languages, but it’s a marked improvement over most other models.

🎉 The Honeymoon Crowd

Developers on X rave that GPT-5 “feels like pair-programming with a senior engineer,” citing the model’s knack for multi-file refactors that compile on first run. Health startups cheer the new safe-completions layer that politely declines uncertain medical advice instead of hallucinating. 

🤨 The Skeptical Chorus

A 600-comment Hacker News thread calls GPT-5 “evolutionary, not revolutionary,” noting most top models now sit within a few percentage points on public leaderboards, and pointing to the steeper $10/M output-token bill for production apps. Privacy advocates remain frustrated that OpenAI still won’t detail training-data sources.

⚖️ Early Verdict

If you’re paying by the token, GPT-5’s larger context window and router efficiency may offset the pricier output. For everyone else, the biggest surprise might be how subtle the upgrade feels: ChatGPT simply answers more often with citations and fewer blunders. Put it through your own workflow (coding, memo writing, or medical Q&A) and decide whether the gains justify the hype.

⚡ Tesla Dojo Shutdown Explained

After years of fanfare, Tesla has disbanded the Dojo supercomputer team. Dojo lead Peter Bannon and about 20 engineers exited for a new startup, DensityAI, prompting Elon Musk to pull the plug and reassign remaining staff to other data-center projects. 

🛑 Why the Project Stalled

DOJO started off strong, but competition, costs, and potentially, lack of attention from executives has led it to an early grave.

• Talent exodus: Multiple chip architects left between 2023-2025, sapping momentum.
  

• Escalating costs: Analysts once projected Dojo could add $500 billion in value, but delays forced Tesla to rely on pricey Nvidia GPUs anyway.
  

• Strategic shift: Musk now touts “Cortex,” a planned external compute cluster, over in-house silicon.

🚗 What It Means for Full Self-Driving

Without Dojo’s custom accelerators, Tesla will likely lean even harder on Nvidia’s H-series chips for training Autopilot—and shareholders may wonder whether the long-promised Robotaxi vision is slipping.

➕ The Deeper Backstory 

Dojo originally chased a jaw-dropping 100 exaflop target, but early prototypes topped out at a fraction of that, all while gulping an estimated 130MW of power—enough to run a midsize data center. Thermal ceilings on Tesla’s wafer-scale D1 chips meant clocks had to be throttled, erasing most cost advantages over off-the-shelf silicon. 

By late 2024 Tesla had quietly invested in Cortex instead, and last month it signed a $16.5 billion deal with Samsung for “AI6” chips destined for cars and robots rather than data halls. Investors may applaud the capital discipline, but surrendering Dojo leaves Tesla’s dream of owning the full AI stack squarely in Nvidia’s hands.

This Week’s Scoop 🍦

🧠 Meta recruits AI moon-shot team “TBD Lab” to chase superintelligence 

💹 SoftBank shares hit record highs after AI-fueled earnings beat 

⛑️ California teams up with AI heavyweights for workforce training 

💻 OpenAI and Nvidia partner on open-weight reasoning models for global devs 

🏛️ White House floats ‘de minimis’ AI rules favoring light-touch oversight 

🚨 Kansas students sue over false AI surveillance alerts in schools 

🎤 Weekly Challenge: Put GPT-5 to Work

Here’s how to do it:

1. 🔩 Nuts-and-Bolts
   Think of a real, nagging glitch: a JavaScript function that won’t fire, an Excel formula that misbehaves, or a paragraph in your legal brief that refuses to read smoothly. Feed it to GPT-5 in fast mode, then thinking mode. Did the slow pass spot deeper logic errors or just add fluff?
   

2. 🧨 Creative Spark
   Hand over something personal: a LinkedIn post you’re drafting, a pitch email for a potential client, or even a caption for your weekend photo dump. Switch to the Listener persona for warmth, then the Cynic for edge. Which voice sounds most like you?
   

3. 🖼️ Big-Picture Strategy
   Grab two conflicting articles on a topic you care about—AI regulation, remote work, the best ramen in town. Ask GPT-5 for a one-page brief you might forward to your boss. Check citations, bias, and originality.

Score each output for usefulness and authenticity, then hit reply and tell us where GPT-5 shined or stumbled. We’ll feature the most insightful trials next week!

Want to sponsor Jumble?

Click below ⬇️

GPT-5 is a major leap in some ways and a let down in others, while supercomputers are scrapped just as quickly as they’re built. The world of AI never stops, and we’re here to cover it all. See you next time! 🚀

Stay informed, stay curious, and stay ahead with Jumble!

Zoe from Jumble