- Thoughts by Humans
- Posts
- Thought #10 - Sunshine, Silly Debates, and Serious AI Updates
Thought #10 - Sunshine, Silly Debates, and Serious AI Updates
OpenAI’s vibes, Meta’s moves, and a leopard-saving AI moment
Laura Gemmell
April 30, 2025
Hi lovely humans,
If you’re in the UK, we hope you’re enjoying the sunshine (long may it last). The AI world is as busy as ever, but we’ve got something a bit lighter for you this week.
We’re having a very silly (but extremely important) debate at Taught by Humans - what do you call OpenAI’s chatbot?
What We’ve Been Up To
AI and Data Resources
We’ve been working on rolling out some AI and Data resources on our website. Some highlights this week:
We’ll be adding to this each week going forward - so any requests, let us know!
AI New Releases
OpenAI Have Had an Odd Week
Shopping
No model releases, but they have improved shopping via their web search (honestly we hadn’t even clocked this was a thing, and we’re not super thrilled by this use case) and are calling for merchants to sign up for their products to be searchable.
Personality Problem
There have been a lot of complaints about ChatGPT’s personality in the past week - too friendly and over the top (sycophantic according to their CEO Sam Altman).
the last couple of GPT-4o updates have made the personality too sycophant-y and annoying (even though there are some very good parts of it), and we are working on fixes asap, some today and some this week.
at some point will share our learnings from this, it's been interesting.
— Sam Altman (@sama)
10:49 PM • Apr 27, 2025
A lot of users have taken to Reddit to say they will be cancelling their subscriptions as it seems the company isn’t testing things properly before rolling them out.
Yesterday (29th April) OpenAI released a statement explaining what happened (valuing short term feedback over long term regarding thumbs up and down to responses) and have rolled back to an earlier version.
o3 Performance and Hallucinations
Last week OpenAI rolled out a new reasoning model o3, the model itself has been met with some backlash - with users taking to social media to complain and the performance on third-party testing being lower than the results released by OpenAI (see Epoch benchmarking and related Twitter thread or try the alternative XCancel if you’re not into Twitter anymore).
Results are also claiming that o3 hallucinates 65% of the time (compared to GPT-4o using web search which is less than 10%) - TechCrunch
We’re a little concerned about some of Meta’s policies towards AI.
Currently they aren’t allowing WhatsApp users to turn off MetaAI
Instagram and Facebook users have been emailed to say all public information for anyone over 18 will be used to train AI, but there is an option to opt out
You can try out their chatbot if you’re interested.
It comes with two modes - thinking and non-thinking. We found the thinking mode felt intense and unnatural (although this could be in comparison to the slickness of OpenAI’s interface.
Play around on Hugging Face without needing to sign up or run any code.
This was only released as we were writing this newsletter - so we will play around a bit and report back.
This study is worth a read, so much so we’re going to call out some findings:
79% of conversations on Claude Code involve AI-driven automation, showing a major shift towards autonomous coding.
Front-end languages like JavaScript and HTML are the most common, suggesting front-end development roles may face disruption first.
Startups are leading in adopting AI coding tools like Claude Code, while larger enterprises are lagging behind.
AI News
Mostly positive use cases of AI this week - payments, saving leopards, freeing up NHS appointments.
Duolingo announces it is becoming an AI-first company claiming it is similar than the shift to mobile-first. They will be using AI instead of contractors where possible.
The White House has announced a new executive order to boost AI education for young people, with a focus on early exposure and teacher training (yay)- though details on funding and delivery remain unclear.
Mastercard have released Agent Pay - a pioneering agentic payments technology designed to integrate seamlessly with AI-driven platforms
Early success with AI doctors freeing up waitlists for the NHS
Oscars set rules around AI usage in movies (and a rule that voters have to watch movies…)
UAE uses AI to write a law (claiming this to be a world first)
New Scientist claims Keir Starmer is being advised by AI
AI, cameras and text messages might be able to save rare snow leopards
Also worth a read
Not exactly news but we thought these were useful and worth reading:
MIT researchers have created a periodic table of how 20 classic Machine Learning models are related
LinkedIn AI Poll
Vote in this week’s poll - please!
It’s about how much AI energy consumption is impacting your usage.
Final Thoughts
As always we hope this was helpful!
Feel free to share this with anyone who might find it useful.
Next week, we’ll share some resources on AI use cases and prompting guides.
Laura
Always learning