📸 Picture This: OpenAI Makes Huge Leap With Image Reasoning in New Releases.

🎨 Ever wanted AI to be able to understand and analyze even your most shittiest drawings? Well, with the newest releases of o3 and o4-mini models from OpenAI, they’ll be able to – even if the drawings are low quality.

After Google dominated the AI space last week, OpenAI have clapped back with these new ground-breaking releases, plus there are reported plans from the AI giant to release their own social media network to rival the likes of X – but more on both of these down below.

Strap in, as ever, there’s a lot to cover given it’s only been a week since our last (and first) Big Machines newsletter.

🗞️ What we are covering today…

• OpenAI’s two new AI models for image reasoning

• OpenAI weigh up creating a new social network

• Nvidia chip sanctions against China imposed by USA

• Computers could soon read our fucking minds?!

• UK creating ‘Murder Prediction’ tool

• Google are trying to talk to dolphins… I can’t believe I have just typed that

🔴 Quick Note: There’s a lot to cover this week, and for a better reading experience, we suggest opening this in your browser! Head to the ‘READ IN BROWSER’ tab at the top of this email.

🍾 FREE ENTRY TO OUR INVITE-ONLY AI CHAT ON TELEGRAM…

To celebrate our launch, if you share this newsletter and tag us on X, we will send you access to our invite-only Big Machines Telegram group, which is full of builders, investors, founders and creators.

This is the final offer on this and will be removed starting next week, as the group loses its pazazz if anyone can join.

👁️ 👁️ What you might have missed

• OpenAI new o3 and o4-mini model releases are now capable of “thinking with images”, allowing AI to analyze and then discuss with you your images or drawings, whether it’s from the whiteboard or sketches. Doing homework may have become a whole lot easier.
  

• Elon Musk vs Sam Altman – Round Two, FIGHT! In last week’s newsletter, we reported on how OpenAI countersued Musk amid harassment claims. Now, just a week on, it seems OpenAI are ready to rival Musk in the social media world, with reports claiming that OpenAI are working on their own X-like social network. While it’s still in the early stages, the internal prototype focused on ChatGPT’s image generation, which has a social feed. That won’t go down well with Mr Musk, we’re sure.
  

• While Donald Trump’s bastard tariffs continue to wreak havoc on the world, Nvidia has also been hit hard as a result of their introduction. Earlier this week, shares in Nvidia plunged and were hit by a staggering $5.5 billion in costs due to the US government tightening its export rules on chips to China.
  

• Here’s an idea for a Black Mirror episode: Computers read the minds of humans and decode their thoughts with 80% accuracy. Wait a minute… checks notes. THIS COULD BE REAL!! Apparently, Meta has published two papers showing that thought-to-action could be just around the corner, using “non-invasive” brain-computer interfaces in real time. The lines between fact and fiction are getting blurrier by the day.

• Google just couldn’t let OpenAI have its time in the sun, could they? Once again, The Goog had a HUGE launch week with an all-new Gemini 2.5 Flash, which is absolutely flying when comparing the price-performance ratio to other models. 

• More, shit from Google… this time it is text-to-video Veo2 and video generation from images, using Whisk. We saw the worlds response to OpenAI’s 4o image generation. Again, there is one-upmanship between these two at the front of the AI sector.
  At the moment, if OpenAI goes to Tenerife on Holiday, Google goes to Elevenerife and vice versa. 
  

• In more dystopian news, the UK government is creating a “murder prediction” tool to identify individuals most likely to commit homicide. According to a report by The Guardian, the AI tool would use algorithmic analysis of personal and criminal data to identify potential offenders. It’s been met with ethical and privacy debates over what could create potential biases.

🗣️ Other Titty Bits 

• Gemini 2.5 Pro criticised over its lack of safety details

• Meta blocks Apple Intelligence across its product suite of Facebook, Instagram and WhatsApp. 

• OpenAI lose ANOTHER Catastrophic Risk Officer… make of that what you will 

• Grok gets memory - Now Grok has context of all your previous conversations when giving you answers

• Midjourney gets a cool image editor 

• Figma tell Lovable they can’t use the term ‘Dev Mode’

• Alibaba’s Wan has released another video model, which looks pretty neat too 

🤳 Picture this: OpenAI makes huge leap with image reasoning in new release.

OpenAI unveiled their o3 and 04-mini models this week, with the newest releases enabling AI to understand images you upload, whether it’s from a whiteboard, diagrams or your drawings only your mother is proud of.

The artificial intelligence company has said their new models are capable of “thinking with images”, meaning it’ll be able to analyze what is being shown and then will break it down for you.

So say if Matt Damon’s character in Good Will Hunting simply just took a photo of the economics math problem on the chalkboard and run it through these new models, OpenAI claims the image manipulation capabilities could provide step-by-step explanations of the image to help understand the content and, hopefully, provide the answer. It would destroy what is a great film, but how do you like them apples?

In fact, we tried it out…

The innovation is seeing OpenAI doing its best to stay ahead of the competition in generative AI, especially after Google’s mega Google Cloud Next extravaganza last week, while Anthropic and Elon Musk’s xAI also maintain pace.

This visual and textual reasoning sees AI take another huge step forward in true multimodal AI which not only sees the images, but can think of solutions to solve them. 

🙂‍↕️ Here’s everything we used this week. 

As mentioned above, this week saw the release of many frontier models, and we have been trying these out in real-time to see which are best and for what use cases.

GPT-o3 and GPT-o4-mini from OpenAI dropped this week and were instantly met with a lot of praise. The o-series of models from OpenAI are designed with more complex tasks and reasoning baked in.

The phrase AGI was being thrown around anecdotally by the likes of Tyler Cowen. Bearing in mind, o3 was the model that in December 2024 scored 73% (low tuned) and 88% (high tuned) when tested using ARC-AGI, arguably the most respected and rigorous test for AGI.

We don’t care about any of that shit. We decide what AGI is in this house…

We have AGI at home… - Test one.

As you can see, it understands basic Beano humour even if the joke is derived from an image. If we are to be subservient under lords to ketamine powered AGI, then I at least want them to have a sense of humour.

We have AGI at home - Test two.

I asked o3 to recreate my favourite childhood game of Beehive Bedlam. If you didn’t have Sky between the years 2000 and 2007, then you missed out on Beehive Bedlam, but don't worry, vibe coders, we can resurrect all the dogshit games that we once used to pass the time.

And, to be fair… it worked even if it was a basic rehash on a terribly shitty prompt of “recreate beehive bedlam for me”

Anyway, I have seen enough. If this isn’t AGI, then I don’t know what is…

In all seriousness for a second, I am impressed with o3 thus far, is it AGI? Who the fuck knows at this point.

Other cool products we found helpful this week (we hope you do too): 

• Kling 2.0 for video generation from either an image or text. Bloody fantastic, despite being pretty damn expensive.

• Veo2 text-to-video from Google is also impressive. I’d say Kling 2.0 is definitely ahead here, but on a quality/cost, Veo2 steals it for me.

• Gamma.App is pretty cool if you want an AI-generated presentation from some sales meeting notes. Super helpful for that personal touch when dealign with potential clients. 

• I was back using V0 from Vercel to help redesign the UI/UX for the blocmates website, and I cannot believe how an idiot like me can get my point across using text-to-website design. Totally blown away every time I use V0. 

• Firecrawl has had a release week during which they have released a ton of new agentic tools to help you scrape any website and manipulate the data in useful ways. Basically, you can turn any website into an LLM-ready TXT file.

• VEED has been immense for basically anything video generation or video editing related. Nice to see some UK-based teams doing great work too.  

• DeepWriter - Personal Scientist, author, assistant and basically everything in between. 

🤝 In Partnership with Mira Network.

🤯 Did you know that AI models hallucinate, and around 30% of their outputs could be false?

This creates significant challenges for anyone relying on AI systems for critical tasks. Mira Network has developed a solution to this problem with its trustless verification network that fundamentally transforms AI reliability.

Their decentralized approach creates a new paradigm for AI verification, enabling users to confidently rely on AI outputs across multiple models (DeepSeek R1, GPT4o, and Llama 3. 70B Instruct as of now).

If you are a developer, Mira's Public Testnet is now live at console.mira.network 

The future of AI isn't just about powerful models—it's about verified ones.

Don’t trust; verify.

💸 Financials

The biggest gold rush since the dot-com boom-bubble-bust is seeing the usually rational and sophisticated venture capitalists forget who they have and haven’t invested in. 

• Incident.io has raised $62 million in Series B funding at a $400m valuation. This investment for the London-based startup company shows the growing importance of resilience and rapid incident response in the tech industry.
  

• Meanwhile, Goodfire has announced their $50million Series A and has shared a sneak peak into Ember, their universal neural programming platform. This give programmable access to any AI model’s internal thoughts, the company claims. Menlo Ventures lead the round whilst Anthropic also contributed $1m.
  

• When OpenAI entered into talks to buy AI coding company Windsurf for $3 Billion, many were left wondering, “why not just buy Cursor instead?” It makes sense, given their startup fund is an investor in the company. But as it turns out, OpenAI did in fact reportedly talk with Cursor in 2024 about an acquisition and subsequently failed. 

• Anysphere (Cursor’s parent company) are instead looking to raise capital at around the $10bn value, according to Bloomberg.

  
  Other financial news 

• Synthesia reach $100m in revenue, secures an investment from Adobe 

• NATO acquire Palantir’s advanced AI warfighting system 🙂‍↔️ 

• Figma acquired Modify and also filed IPO papers

🐣 Early Stage Products and Companies

• Plastic Labs raised $5.3m to solve personal identity in AI 

• Maskara AI are building AI that builds AI.

• OpenPhone released Sona, an AI voice agent who is always on call for your company. 

• Kular releases a lead generation agent (first lead is free, apparently, too!)

• Voice-based conversational interviews by an agent are finally here with Light Screen.

If you are a new company and produce that would like to be featured here please shoot us a dm on X!

📆 The week ahead

• Dubai AI Week will see over 10,000 AI experts, CEOs and tech visionaries head to the United Arab Emirates next week, with delegates from OpenAI, Google, Meta and Microsoft take part in panels discussing ethics in AI along with providing live demos of generative AI tools.
  

• We are also half expecting a big raise announcement from Mira Murati and her new company, Thinking Machines, but we will see. If we are to guess, it would be close to the $30bn raise we saw from her fellow OpenAI alumnus, Ilya Sutskever, and his new venture SSI. 

👋 Until next week

Now, here are two of the best pieces of AI-related news/content this week for you to enjoy this fine weekend. 

First, Google are literally trying to talk to dolphins… no, we are not lying, they are trying to communicate with the aquatic bastards. 

I mean, if this gets us closer to being able to understand our pets, then have at it!

The problem is that a lot of people who have done questionable things around the house in front of their pets might want to start being a little nicer to them if they are about to have the ability to communicate… 

While there were plenty of phallic references in our first newsletter, it seems we’re turning the other cheek this time round. Velvetshark.com has hilariously spotted that all AI company logos all look like buttholes. They have a point, to be fair – have a look for yourself:

What do you think of the news discussed above? Think the murder prediction tool is something that could work, or is it utterly mad? 

And as always, feedback, good and bad, is always welcome, so drop us a line on what you like and don’t like, and we’ll ignore all the negative ones sent in. 

If you are feeling really generous, please share this with your nerdy mates in the group chat to help us grow, so we can pay poor Sam, or we may have to replace him with AI.

See you next week,

Sam, Grant and The Big Machines team.