Cheap thinking & more (April 18, 2025)

Hey friends,

It’s getting weird out there.

Google just launched a budget model that lets you toggle how much it thinks. OpenAI’s o3 is sparking AGI rumors (again). And someone built what’s basically a USB stick for your chatbot memories. We’re officially in the “AI gets weird and powerful at the same time” era.

Let’s dive in.

⚙️ MODEL WARS

Google launches Gemini 2.5 Flash with a "thinking budget"
Google’s new Gemini 2.5 Flash model lets you toggle thinking on/off — literally. You can dial reasoning up or down based on cost, speed, or quality, and it outperforms Claude 3.5 Sonnet on STEM tests while matching OpenAI’s o4-mini.

Thinking budget = customizable cost-efficiency (up to 24k tokens)
Available via Google AI Studio, Vertex AI, and the Gemini app
Visual reasoning and hybrid logic packed in at a lower price

OpenAI’s o3 sparks fresh AGI discourse
o3 can now use tools and dominates benchmark charts. Tyler Cowen calls it smarter than people. But others point to its weird fails on basic logic puzzles. One thing’s clear: AGI might arrive in layers, not lightning bolts.

Passed the Turing Test 73% of the time
Experts are split: genius-level or just clever marketing?
Meanwhile, GPT-4.5 is quietly climbing up behind it

Meta’s FAIR drops 5 open-source AI tools
Meta’s research arm is on fire: visual encoders, 3D spatial understanding, and a “collaborative reasoner” that performs 30% better than solo models.

🔬 WHAT’S COOKING

Profluent proves scaling laws in protein design
Its ProGen3 model designed proteins from scratch, including gene editors smaller than CRISPR — and new antibodies that don’t violate patents.

Trained on 3.4B sequences
Offering 20 open antibody designs royalty-free
Could turn drug dev into software-level iteration

Google Sheets gets prompt-native AI functions
New formula: =AI("Summarize this in 3 bullet points", A2)
It works like magic. Combine it with formulas like IF() or drag it across rows for batch processing. Also: you can prompt formatting instructions right in the cell.

Goodfire raises $50M to steer model behavior
Interpretability startup Goodfire is building tools to control how AI models think. They specialize in steering, SAEs, and aligning motivations with intent.

⚒️ TOOLS WE’RE TESTING

Codex CLI – OpenAI’s new terminal-based coding agent
Gamma 2.0 – Generate websites, decks, and socials from text
Webdraw – Remix and build full AI apps, no code needed
Docwelo – Auto-create NDAs, contracts, and client docs in minutes

💬 PROMPT OF THE DAY

"Prepare me for an interview."
Prompt: Act as a senior hiring manager in [industry]. Ask me 8–10 hard interview questions. Grade each response and tell me how to improve. Then rewrite my ‘Tell me about yourself’ answer to be high-impact and role-specific.

⚡ QUICK HITS

o3 scored a 136 on Mensa Norway’s IQ test — highest ever
DeepSeek may get banned from U.S. chips (and platforms)
LMArena breaks out into a benchmarking company: Arena Intelligence
xAI’s Grok now remembers conversations and organizes files
Perplexity is in talks with Motorola and Samsung
YouTube’s Dream Track now auto-generates soundtracks for Shorts

TL;DR

Google’s newest model lets you toggle AI reasoning like a thermostat. OpenAI’s o3 might be inching toward AGI (or at least looking the part). Meanwhile, spreadsheets talk back, proteins become programmable, and the AI memory wars have officially begun.

🧠 Until next time...

Whether you’re toggling your own thinking today or letting an LLM do the heavy lifting — keep an eye on the edge. These updates feel small now, but they’re setting up some massive shifts.

Catch you in the latent space,
David