2025-11-13: Gemini-3
Some more solid early details
🔷 Subscribe to get breakdowns of the most important developments in AI in your inbox every morning.
Here’s today at a glance:
🧪 Google’s Testing Leaks
Gemini-3, the next major release for Google, has been in testing in Google’s AI Studio for a while. While various rumors about capabilities have circulated, I finally have enough confidence to write about it after reading a stellar report by historian Mark Humphries on his blog at Generative History. And so here is an early, rumor-filled, unconfirmed report of Gemini-3’s capabilities. These were collated from X posts from various anonymous machine learning community members during Google’s A/B testing in LM Arena and Google’s AI studio in October.
📜 Generative History
According to Mark Humphries, Gemini 3 achieved 2 things:
handwriting recognition of historical texts
inferred meaning of such texts
He uses the example of a ledger kept by the Dutch speaking clerk of a New York merchant in 1758 (18 years before the American revolution). It is a mess of multiple languages, non standard spelling, abbreviations, mistakes, corrections, terms of art, and calculations, with multiple non decimal currencies.
In tabulating the “errors” I saw the most astounding result I have ever seen from an LLM, one that made the hair stand up on the back of my neck. Reading through the text, I saw that Gemini had transcribed a line as “To 1 loff Sugar 14 lb 5 oz @ 1/4 0 19 1”. If you look at the actual document, you’ll see that what is actually written on that line is the following: “To 1 loff Sugar 145 @ 1/4 0 19 1”. For those unaware, in the 18th century sugar was sold in a hardened, conical form and Mr. Slitt was a storekeeper buying sugar in bulk to sell. At first glance, this appears to be a hallucinatory error: the model was told to transcribe the text exactly as written but it inserted 14 lb 5 oz which is not in the document. This was exactly the type of errors I’ve seen many times before: in the absence of good context the model guessed, inserting a hallucination. But then I realized that it had actually done some extremely clever.What Gemini did was to correctly infer that the digits 1, 4, 5 were units of measurement describing the total weight of sugar purchased. This was not an obvious conclusion to draw, though, from the document itself. All the other nineteen entries clearly specify total units of purchase up front: 30 gallons, 17 yds, 1 barrel and so on. The sugar loaf entry does this too (1 loaf is written at the start of the entry) and it is the only one that lists a number at the end of the description. There is a tiny mark above the 1 which may also (ambiguously) have been used to indicate pounds (thanks to Thomas Wein for noticing this). But if Gemini interpreted it this way, it would also have read the phrase as something like 1 lb 45 or 145 lb, given the placement of the mark above the 1. It was also able to glean from the text that sugar was being sold at 1 shilling and 4 pence per something, and inferred that this something was pounds.To determine the correct obverse weight, decoding the 145, Gemini then did something remarkable: it worked through the numbers, using the final total cost of 0/19/1 to work backwards to determine the weight, a series of operations that would require it to convert between two decimalized and two non-decimalized systems of measurement.Generative History
Humphries points out, specifically, that this is far away from predict-the-next-token as possible. The texts are non-standard. One has to immerse oneself in the milieu and learn an almost code like language that only has specific meaning in the particular context it is used in.
This for me was the final confirmation needed to announce that Gemini 3 is likely to be a landmark release. A fully multimodal reasoner of astonishing capability.
In any case, a round up of less confirmed capabilities evinced in the last month.
One API. End-to-end Voice AI Stack
Everything you need to build intelligent Voice AI, all in one API. AssemblyAI combines transcription, speaker identification, PII redaction, and LLM integration into a single, production-ready platform. No need for five separate tools, just one API to build, test, and scale powerful voice products fast.
Try it free.
📇 It makes a working OS
Windows 95
Mac
Playstation
📲 Builds websites, of course
with high design
🎶 It composes music
with visualizations
👾 It makes games
📤 Comparing 1 year ago
For reference here is a Llama 3.1 70B pelican SVG generated by Simon Williamson last year
🍎 Apple
This now explains the Apple rumor:
Which means Google has previewed a Gemini 3 checkpoint to Apple. It’s a 1.2 trillion param model, and Apple is signing on. And they must be productization mode right now.
This would resolve Apple’s biggest issue right now, and help Google cutoff OpenAi’s prime distribution channel.
We are now in a bloody knife fight to the finish. Stay tuned.
🖼️ AI Artwork Of The Day
💻 One More Thing
We have sponsors (or at least one sponsor)! If you'd like to explore partnerships with Emergent Behavior, email ai@a16zstudios.com.