The Neuron
Posts
😺 DeepSeek Returns...

😺 DeepSeek Returns...

PLUS: ChatGPT Search has Shopping now?!

Grant Harvey
April 29, 2025

Welcome, humans.

Users have been complaining over the last week about OpenAI’s latest 4o update and its new personality that treats every prompt like it's Nobel Prize-worthy, with some describing it as manipulative and counterproductive.

(fun tip: here’s a prompt to get the complete opposite of 4o’s current shtick, and here’s another, cruder option).

Sam Altman even admitted as much this weekend, saying the updates have made the AI “too sycophant-y and annoying” (lol) and promised fixes are coming ASAP.

The company is not only working on immediate fixes, but Altman hinted we might eventually get to choose between different AI personalities—fingers crossed for a “slightly grumpy but still gets your ish done” option!

In other news, here’s what happens if you ask GPT’s imagen to “create the exact replica of this image, don’t change a thing” 74 times.

Spoiler: it gets * ~WeirD dUdE ~ *.

Here’s what you need to know about AI today:

DeepSeek R2 was leaked over the weekend.
White House ordered agencies to scale up AI use.
OpenAI rolled out ad-free shopping in ChatGPT.
IBM launched ATOM and PTI to automate security operations.

DeepSeek R2 could crush AI economics with 97% lower costs than GPT-4

Remember when DeepSeek R1 shocked the market and wiped out $1 trillion in stock value? That was just the warm-up act.

Rumors flew over the weekend that the Chinese AI startup DeepSeek is accelerating the release of its next-generation model, DeepSeek R2.

Originally scheduled for May 2025, sources indicate it could drop as early as April (so like, Wednesday) due to competitive pressure and market momentum.

The leaked specs are mind-blowing:

A massive 1.2 trillion-parameter model using a hybrid MoE (Mixture-of-Experts) architecture.
Only needs 78 billion active parameters per token (that's efficiency!).
97.3% cheaper to train and run than GPT-4.
Expected pricing of just $0.07 per million input tokens and $0.27 per million output tokens.
Trained on 5.2 petabytes of data across various domains.

In a single sentence, R2 promises enhanced coding capabilities across 30+ languages, multilingual reasoning, advanced multimodal support, and potentially a massive 128K token context window.

The most industry-disrupting aspect? Analysts estimate DeepSeek's pricing could be 20 to 40x cheaper than what you're paying for GPT-4o right now.

What makes DeepSeek truly revolutionary is their complete tech independence—#1, using Huawei Ascend chips instead of NVIDIA hardware, and #2, innovative architecture that only activates necessary model parts for each task.

Also, these guys have work life balance(ish). Unlike typical Chinese tech companies with brutal “996” work cultures, DeepSeek maintains a collaborative environment where employees are “treated as experts.” This approach has paid off—their models have outperformed Meta's Llama 3.1 and OpenAI's GPT-4o at a fraction of the cost.

For fun: this is what ChatGPT thinks DeepSeek R2 looks like lol.

Why this matters: The implications are enormous. If a player can deliver comparable performance to today’s top AI at 3% of the cost, the entire industry could change. As one tech executive put it, DeepSeek's approach could break “the stranglehold of the few dominant players in the field.”

Remember, the original DeepSeek R1 could do a ton of wild stuff. If this new version can do substantially more, China might actually have a frontier AI model on its hands. Be prepared for it to be banned in the US almost as quickly as it arrives.

FROM OUR PARTNERS

🧰 The Tools, Templates & Playbook for Your AI Consultancy

Remember when AI experts were just “that person who knows ChatGPT”?

Well, companies are now dropping serious cash—the AI consulting market is exploding 8x from $6.9B to $54.7B by 2032.

The problem? Everyone wants to be an AI consultant, but most are just winging it. No system, no frameworks. Translation? They're leaving money on the table while working twice as hard.

Our friends at Innovating with AI have welcomed 700 students into The AI Consultancy Project.

Here’s why they like it:

Tools to find clients and deliver top-notch services.
A 6-month plan for a 6-figure consulting business.
First clients in as little as 3 days.

Click here to request access

Prompt Tip of the Day

Heads up: apparently “em dashes” have been flagged as an “AI tell.” Don't panic—they're actually marks of good writing! As Joel Dueck explained “The AI has learned this practice from writers and editors who are good at what they do, and you should follow its example.”

Longtime Neuron readers may have noticed we're guilty of a lil’ dash-love ourselves—we often add ‘em in AFTER writing our drafts—because they create breathing room that commas just can't. As typography expert Matthew Butterick says, good typography “helps your reader devote less attention to mechanics and more to your message.” So basically we do it for you… you’re welcome.

Instead of stripping em dashes to hide AI use, try adding this sentence to your next writing prompt: “Use em dashes where they genuinely improve rhythm—not excessively, but purposefully.”

P.S: yesterday we shared one idea from Reddit of someone uploading their lab results to AI to explain it all to them. Here’s a prompt to do that without jeopardizing your privacy (after you remove your personally identifying details). Don’t feed the AI more than you need to!

Treats To Try.

*Data brokers selling your personal info? Yikes. Incogni Unlimited not only scrubs your data from our extensive 250+ broker network, but lets you point them at ANY sketchy site sharing your details. Protect your personal data for 55% off with code NEURON
Alibaba released Qwen3, which upgrades their open source model family with dual thinking modes, 119 language support, and better performance across eight different model sizes trained on double the previous data—free to try.
Promptimize enhances your prompts with one click across all AI platforms, giving you expert-level results without prompt engineering skills—free plan, then $12/month.
Tufa creates and schedules social media posts tailored to your industry—free trial, then $8.99/month.
Hera turns your text into stunning motion graphics instantly—free trial available.
ExcelMatic answers questions about your spreadsheets, giving you instant charts and insights—free to try.
Pikzels generates viral YouTube thumbnails from text prompts in 30 seconds—free trial, then plans from $14/month.
Not sure if we should include this, might be TOO controversial, BUT… Dog-e-dex helps you snap, save, and categorize dogs you meet—free to try.

See our top 51 AI Tools for Business here!

*This is sponsored content. Advertise in The Neuron here.

Around the Horn.

The US White House directed all federal agencies to rapidly scale up their use of AI, including for healthcare diagnostics and insurance eligibility.
OpenAI rolled out new updates to ChatGPT search, incluing ad-free, personalized shopping features with images, reviews, and direct purchase links.
- Also, OpenAI says ppl made 1B+ web searches last week and if its your jam, you can text 1-800-CHATGPT for web results.
IBM introduced ATOM and PTI, agentic AI systems automating threat detection, analysis, and response across multiple intelligence sources—oh yeah, it’ll also invest $150B in the US over the next five years.
A University of Melbourne-KPMG survey of over 48K people found that three in five respondents in emerging economies trust AI (versus two in five in advanced nations), with 83% seeing benefits despite 58% expressing distrust.
The US House of Representatives passed a bipartisan bill to criminalize non-consensual deepfake adult material and requires platforms to remove such material.