In partnership with

🔧 Sam Altman Confronts GPT-5 Backlash

OpenAI’s CEO is publicly owning the bumpy GPT-5 debut. At a press dinner in San Francisco, he said the team “screwed that up,” and detailed steps to steady the product. The conversation pointed beyond model scores to a wider strategy across apps, hardware, and even a possible browser.

OpenAI also acknowledged the rocky first week in a note on the bumpy rollout that previewed quick product changes. For context on official specs, see the GPT-5 announcement.

🧨 What Went Wrong

Users balked at a colder tone, the automatic router, and the removal of choices they relied on. In response, OpenAI restored access to GPT-4o and returned model selection controls, which now include Auto, Fast, and Thinking modes with clearer limits. Details are in a report on the model picker returning and a restoration update on GPT-4o access.

🛠️ Fixes in Flight

OpenAI is shipping updates that make GPT-5 warmer without sliding back into flattery. The company also promised future “transition periods” before deprecating models and said it is testing a rubric with mental health experts to push back on unhealthy use.

Early signals suggest demand is not the problem: Altman said API traffic doubled within 48 hours and that the company is “out of GPUs,” as covered in a follow up dinner piece. A separate note says the team is “making GPT-5 nicer” with targeted tuning, captured in an update on tone improvements.

🚀 Altman’s Bigger Bet

OpenAI hired Fidji Simo as CEO of Applications to build consumer products beyond ChatGPT, and Altman teased unannounced apps, a potential AI browser, and long term bets from devices with Jony Ive to brain computer interfaces. Reporting on the latter appears in a piece on OpenAI’s talks to back Merge Labs. The message is clear: future headlines will be as much about experiences and distribution as raw model lifts.

Former Zillow exec targets $1.3T

The top companies target big markets. Like Nvidia growing ~200% in 2024 on AI’s $214B tailwind. That’s why the same VCs behind Uber and Venmo also backed Pacaso. Created by a former Zillow exec, Pacaso’s co-ownership tech transforms a $1.3 trillion market. With $110M+ in gross profit to date, Pacaso just reserved the Nasdaq ticker PCSO.

Invest for $2.90/Share

_{Paid advertisement for Pacaso’s Regulation A offering. Read the offering circular at invest.pacaso.com. Reserving a ticker symbol is not a guarantee that the company will go public. Listing on the NASDAQ is subject to approvals.}

🗺️ Google Translate  ‘Practice Mode’ Targets Duolingo

Google is trialing a Practice mode inside Translate that behaves like a lightweight tutor rather than a simple dictionary. Early builds surfaced by an APK teardown point to personalized lesson plans, scenario prompts, daily activities, and first-language support for French and Spanish.

Screens and feature behavior are detailed in a hands-on look at Practice mode. The feature sits alongside core translation, so it does not replace the main app’s camera, offline packs, or phrasebook.

🧪 What Practice Mode Does

Practice asks learners to choose a starting level, then runs short rounds that mix listening, pronunciation, and speaking drills. You can generate your own scenarios by prompt, for example “order coffee in Paris” or “check into a hotel in Mexico City,” and the app adapts difficulty as you succeed. A goals view tracks words reviewed and streak-style progress. 

Because it lives inside Translate, Practice can also draw on live translations to seed quick sessions around words you already look up. In the builds reviewed so far, progress appears device-local and there is no subscription upsell, which positions Practice as a light, free coach rather than a full curriculum.

🗣️ Duolingo’s Reframing and the Competitive Angle

Duolingo’s CEO Luis von Ahn is clarifying a widely criticized “AI-first” memo, telling the Times he “did not give enough context” and that full time roles are safe. Recent coverage sums up his comments and the user backlash in a recap of the CEO’s interview and memo. 

If Practice rolls out widely, Google will compete not only on translation but on habit loops that Duolingo has perfected. Expect Duolingo to lean on engagement mechanics, brand voice, and premium features while Google tests how far light coaching can go inside a utility app.

This Week’s Scoop 🍦

🃏 Grok persona prompts leak exposes “conspiracist” and “unhinged comedian” modes

📉 Five year tech forecast flags surprising winners and losers

🫨 ‘AI psychosis’ is happening at an alarming rate

🐉 Nvidia reportedly prepares a new China bound chip that tops H20

🛡️ Anthropic tightens its usage policy for a riskier AI landscape

🧪 Google Labs expands Flow Whisk and adjusts AI credits

💪 Weekly Challenge: Discover 10 GPT-5 Power Prompts and Modes You Should Try

1) 🔺 Pick the right mode for the job
Use Auto for most chats. Use Fast for quick drafts. Use Thinking for multi-step reasoning. Mode details and the restored picker are in a walkthrough of the model switcher.

Try: “Solve this exactly. Show numbered steps. Use no outside facts: 3814 × 207. Then verify by reversing the operation.”

2) 📝 Push long context hard
GPT-5 supports very large inputs; OpenAI highlights longer documents and codebases in the product introduction.

Try: “Read this 120-page policy PDF. Return a two page brief with a risk register, five unresolved questions, and a one paragraph CEO summary.”

3) ✍️ Make it write, then make it prove it
Try: “Draft a 300 word memo that argues for a travel budget. After the draft, list three citations with links that support each claim. Label any claim without a source as unverified.”

4) 👯 Use contrastive pairs to tune tone
The rollout included work on warmth and style, covered in an update on tone improvements.

Try: “Rewrite this email twice: Version A stoic and concise. Version B warm and collaborative. Add a short rubric telling me when to pick each.”

5) 🧑‍🎓 Let the model build and grade its own rubric
GPT-5 is stronger when you give it criteria and it to self-check.

Try: “Before you answer, write a five item rubric for great product specs. Answer the prompt. Grade your answer against the rubric. Fix any misses.”

6) 🗣️ Ask for a plan you can execute in steps
Try: “Return a step list to refactor this codebase. After the list, ask me to paste one file at a time. Wait for my signal ‘next file’ before continuing.”

7) 🧑‍💻 Force crisp outputs with schemas
It follows schemas incredibly well.

Try: “Return JSON only with keys: goal, steps[], risks[], time_estimate_minutes, sources[]. Do not add prose.”

8) 🧑‍🏫 Do an A/B mode test on the same task
The picker came back after user feedback, as reported in a note on restoring GPT-4o and model choice.

Try: “Answer in Fast and in Thinking on the same question. Time each response. Tell me which is better and why in three lines.”

9) 🔍 Use retrieval-style constraints even without tools
Try: “Answer only with information that appears in the pasted text. If a detail is missing, say ‘unknown in source’.”

10) 🏁 Code with checkpoints
OpenAI positions GPT-5 as a stronger coding collaborator in the developer overview.

Try: “Given this repo, propose a three commit plan with commit messages. After each commit, run tests conceptually and report expected outcomes. Stop if any test would fail.”

Want to sponsor Jumble?

Click below ⬇️

What change would have made the GPT-5 rollout feel smoother on day one for you? Would you use AI to help you learn a new language? See you next time! 🚀

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

Zoe from Jumble