Welcome, humans.

Armchair researchers (a.k.a Redditors) are now stress-testing models with complex, multi-object requests like “generate 100 animals with their names” or “create 40 movie posters with titles” (just don’t look too closely at the Shrek one… trust us).

The results = A beautiful mess of garbled text (“SNAKA” the snake, “Dart Gump”, etc), morphed hybrid creatures like giraffe-rhino hybrids, and names that just... give up halfway through.

These kind of complicated infographic images are basically the new benchmark for image models IMO. They’re testing the visual equivalent of context rot: that phenomenon where AI models lose track of details the longer a conversation goes on.

Let’s call it “composition collapse”: the more objects you ask for, the more the model's brain melts, producing increasingly cursed creations that somehow get funnier the longer you scroll. Actually maybe we call it AI brain rot?

Here’s what happened in AI today:

• Anthropic showed Claude completed months-long genomic research.

• OpenAI signed a $10B deal with Cerebras Systems for fast compute.

• Higgsfield became the new “fastest growing” AI startup.

• An eighth wrongful death lawsuit alleged ChatGPT acted as a "suicide coach"

Two shameless plugs!

1. Later today, we’re going LIVE at 9am PST | 12pm EST | 3pm GMT to talk AI’s impact in industry, and we’ll be joined by two special guests: Joe Salesky of Recon Analytics and our very own O.G. co-founder of The Neuron, Noah Edelman!

2. We had the pleasure of joining our friend Logan Lawler on Dell’s Reshaping Workflows podcast to talk about the future of AI and how we use it to write The Neuron; Grant also got to tell Logan how he crashed a Dell machine; fun times! (watch on YouTube; also check out Logan on our podcast ICYMI!).

Claude Is Becoming Your Researcher’s Favorite Lab Partner

If you've ever asked ChatGPT or Claude a science question, you've probably gotten a decent summary. But what if AI could actually do the tedious parts of research; the weeks spent cleaning data, the months combing through literature, the guesswork about which experiments to run?

Well, Anthropic just published some serious case studies showing that's already happening in some of the world's top labs.

Here's what researchers are building with Claude:

• Biomni (Stanford): A biomedical AI agent that connects Claude to hundreds of databases and tools.

  • One test = a genome-wide association study (normally a months-long project of data cleaning and analysis) completed in 20 minutes.

  • Another analysis that would take a human expert three weeks finished in 35 minutes.

• MozzareLLM (Cheeseman Lab, MIT): After CRISPR knocks out genes in millions of cells, someone has to interpret what the patterns mean.

  • Lab director Iain Cheeseman can recall about 5,000 gene functions from memory, but the AI still catches connections he misses.

• Lundberg Lab (Stanford): Instead of asking “what do we already know about this disease?”, they're testing whether Claude can predict which genes matter based on molecular properties.

  • This could potentially beat the old method of grad students arguing over a Google spreadsheet.

The bigger picture: Anthropic's AI for Science program gives free API credits to researchers working on high-impact projects, particularly in biology. And Claude for Life Sciences now connects directly to tools like Benchling (lab notebooks), PubMed (literature), and 10x Genomics (single-cell analysis).

This idea with all this is to eliminate the bottlenecks that keep brilliant people stuck on data cleanup instead of discovery. As one researcher put it, the usefulness keeps growing with each new model release. If you're in research and want API credits, applications are open.

And in case you’re wondering why this is all such a big deal to Anthropic, you can read CEO Dario Amodei’s Machines of Loving Grace. As with all great pursuits… it’s personal.

FROM OUR PARTNERS

Introducing the Voyage 4 series

MongoDB’s Voyage 4 series is a unified family of text embedding models built for flexible, cost-effective context-engineered agents and enterprise search.

The Voyage 4 series introduces shared embedding spaces designed to make it easier to switch between models in the same family without re-indexing. You can mix and match models for documents and queries within a single pipeline:

• voyage-4-nano for local development

• voyage-4-lite for query workloads

• voyage-4-large for production indexing, using Mixture-of-Experts architecture to deliver state-of-the-art accuracy with up to 40% lower serving costs

Explore the Voyage 4 series and see how it fits into your embedding pipeline.

Prompt Tip of the Day

Marketers, this one’s for you: there’s a new list of “cursed” phrases that you should probably avoid because it gives away your content as “botspeak.”

Here’s the TL;DR of the phrases to avoid (you might want to copy + paste that whole thread into a chat window and have GPT or your fave bot put all the phrases in a short list you can add to a “ban” list):

From the OP:

1. "And honestly?" — Unnecessary sentence starter before saying something not particularly honest

2. Therapist speak — "You're not imagining it/alone/broken/weak"

3. Forced depth — "Do you want to sit with that?" / "Are you ready to go deeper?"

4. Hype phrases — "Here's the kicker" / "And the best part?" / "Here's the part most people miss"

5. Verbose signposting — "I'm going to state this clearly" + 600 words that could be 2 sentences

6. "Here's the breakdown:"

7. Everything "quiet" — "quiet truth," "quiet confidence," "quietly growing," "quiet rebellion"

8. Forced validation — "You're right to push back on that"

Top-Voted from Comments:

• "That's rare” — The new tell everyone's noticing

• "And that matters" — Appears in nearly every response

• "It's not X, it's Y" — The classic contrast framing

• "Let's unpack this" — Overused transition phrase

• Small caption formatting — Breaking text into tiny labeled sections

• "You're allowed to...” — Unsolicited permission-giving

• Excessive emoji bullets — ✅ Organizing ✨ Everything ⭐ With 🎯 Emojis

• "Let's sanity-check this"

• Awkward metaphors — Comparisons that sound smart but don't quite work

The consensus: these phrases now instantly flag AI-generated content to anyone paying attention. You should probably drop ‘em.

Treats to Try

1. *Scroll.ai turns any knowledge base into an enterprise-grade chatbot or agent. Scroll delivers accuracy, nuance, and depth that generic models can't touch. Use the NEURON2026 coupon for 2 free months of Starter ($158 value).

2. FLUX.2-klein (4B version) generates or edits images from text prompts in under 1 second through ComfyUI—non-commercial license.

3. TranslateGemma translates across 55 languages and runs on your laptop or desktop—handles both text and images, available in 27B, 12B, and 4B sizes based on your hardware—free to try.

4. ChatGPT Translate is OpenAI's first standalone utility tool competing directly with Google Translate, while TranslateGemma democratizes translation by making state-of-the-art models runnable on consumer hardware.

   1. Use ChatGPT Translate when: you need contextual understanding, tone adjustments, or want to ask follow-up questions.

      Use TranslateGemma when: you need offline translation, want complete privacy, or have limited/no internet access.

   2. Note: you’ll be able to use TranslateGemma via LM Studio, but that model might not be ready yet, so until then, you can use this Tencent Model.

5. Anthropic added MCP Tool Search to Claude Code, enabling “lazy loading” of tools that reduced context usage from 134K to 5K tokens.

6. Clawdbot runs locally and automates your entire workflow through WhatsApp/Telegram/Slack/Discord—controls Chrome browsers, manages email, runs cron jobs, handles flight check-ins, controls IoT devices, and replies autonomously using your own Claude/GPT API keys (docs)—free and open source (you just need your own API keys).

7. OpenWork is an open-source desktop GUI for OpenCode—lets non-technical users automate workflows like controlling Home Assistant and deploying apps without terminals, with installable skills via package manager—free to try (MIT).

8. Gambit is an open-source agent harness that inverts pipelines—agents call agents in parallel instead of orchestrated compute, define interfaces in markdown/TypeScript, and auto-grade every turn to catch PII leaks—free to try.

9. 1Code gives you a Cursor-like UI for Claude Code—run agents in background worktrees without touching main, preview diffs before they land, manage terminals, and switch workspaces with Ctrl+Tab (site)—free and open source.

P.S: Your procurement platform needs a rebuild—which means you can finally ditch the legacy system. Coupa and Solenis will walk through modern spend management options on January 28 at 11:00 AM ET. See what's actually possible now.** 

**For transparency, this is a partner event from our parent company, not a sponsor!

Around the Horn

1. OpenAI signed a multi-year deal with AI chipmaker Cerebras Systems worth over $10B to deliver 750 megawatts of ultra low-latency compute capacity through 2028, aiming to speed up AI inference for faster model responses.

2. Symbolic.ai, an AI journalism startup founded by former eBay CEO Devin Wenig and Ars Technica co-founder Jon Stokes, signed a deal with News Corp to deploy its AI publishing platform starting with Dow Jones Newswires.

   1. Symbolic claims productivity gains of up to 90% for complex research tasks in areas like newsletter creation, fact-checking, and SEO optimization.

3. Taiwan agreed to invest $250B in US semiconductor manufacturing through 2028 alongside $250B in credit guarantees as part of a Trump administration trade deal; also, it sounds like Apple is struggling to secure chip production capacity as NVIDIA likely overtook it as TSMC's largest customer in at least one or two quarters of 2025.

4. OpenAI’s former head of its safety research team focused on mental health responses (Andrea Vallone) joined Anthropic's alignment team under researcher Jan Leike, who himself departed OpenAI in 2024 over safety concerns.

5. Higgsfield raised $80 million at a $1.3 billion valuation while reaching $200 million ARR that doubled in two months, claiming to be the fastest-scaling GenAI company in history.

6. Self harm story: A lawsuit filed by Stephanie Gray alleges ChatGPT acted as a “suicide coach” for her 40-year-old son Austin Gordon, transforming his favorite childhood book “Goodnight Moon” into a “suicide lullaby” before he died by suicide in November 2025, marking at least the eighth wrongful death lawsuit claiming ChatGPT contributed to users taking their own lives.

Intelligent Insights

In today's Intelligent Insights*, we cover AI Explained's breakdown of Claude Cowork's surprising failures, DeepSeek's Engram paper on language model’s wasted resources, Alberto Romero's 8-hour beginner-friendly AI tutorial, and a few pieces debating the question about what the economy will do with infinite custom software…

*Why is this on the website? Because email clients cut us off if we go too long, and we want you to have the best resources possible as soon as possible!

A Cat’s Commentary

P.P.S: Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here.