AI
Intelligence Just Got Cheaper at Every Layer, From a 722 Paper Math Dump to Sub One Bit Models
Intelligence is having a deflationary decade, and this week compressed about a year of it into seven days. On Monday a single OpenAI model released more mathematical research in one GitHub commit than most departments produce in a career. Days later OpenAI priced near frontier intelligence at one fifth of its flagship rate. And Samsung researchers kept shrinking the models themselves, squeezing capable language models into less than one bit per weight. Three different teams, three different layers, one direction. Thinking is getting cheaper to produce, cheaper to buy, and cheaper to carry in your pocket.
The math drop landed on October 6, 722 manuscripts organized into 372 result families, every one generated by an unreleased internal OpenAI model that had been posed roughly 4,000 open problems. OpenAI says each surviving result consumed about three hours of ChatGPT Pro equivalent compute on average, with a share of the proofs machine checked in Lean. Within a day the repository showed its first corrections, three manuscripts withdrawn over a sign error that invalidated a shared construction, fourteen more revised with proof repairs. The mathematical establishment split down the middle. The Association for Human Mathematics called the release a demonstration of power and urged mathematicians to step back from collaborating with the company, while Anthropic mathematician Levent Alpoge called it the most significant moment in mathematical history for what it shows about AI's usefulness to working mathematicians. Terence Tao republished the association's statement on his own blog and argued the field now has to rethink what solving a problem even means when producing solutions has become the easy part.
Then came the price cut that reframes what frontier AI costs. At its DevDay on September 29 OpenAI launched GPT 6.1 Sol at $2 per million input tokens, $0.10 for cached input, and $10 per million output tokens, exactly one fifth of the GPT 6 Astra rate card while OpenAI positions Sol close to Astra on coding, computer use and professional work. The cached input number does the heaviest lifting for developers, at $0.10 it makes long agent loops with repeated context dramatically more affordable. The event also rolled out always on Dots agents, a Codex cloud for coding work, an Ultrafast speed tier and a Pro 500 plan, but pricing was the headline the whole industry felt. When near frontier capability sells for a fifth of the price, projects that stayed stuck in a budget review six months ago get a second life.
The third story is the quietest and maybe the most consequential for everyday devices. Samsung Labs released LittleBit 2, a compression technique that squeezes large language models into a range of 0.1 to 1.0 bits per weight, well under the one bit floor that used to define the extreme end of quantization. At 0.1 bits per weight LittleBit beats competing methods running at seven times the storage on Llama2 7B, compresses Llama2 13B to under 0.9 gigabytes for a roughly 31 times memory reduction, and unlocks up to an 11.6 times inference speedup over FP16, with the code public on GitHub. It arrives as the industry argues over chip supply and memory prices, a reminder that you can shrink the model instead of growing the hardware. The other half of the AI cost equation just got a serious upgrade.
Put the three stories together and the pattern glows. A model that writes 722 math papers in a week, frontier intelligence at a fifth of the price, and serious models running in under a gigabyte. The cost of a good thought keeps collapsing, and every collapse moves power from the owners of giant data centers to the builders with a laptop and a sharp idea. For readers that means the same thing it has meant every cycle, learn the tools while they are cheap, build while the giants are busy arguing about supply, and remember the advantage always lived in the clarity of the mind using it.
Quick answers
What is this story about?
Intelligence is having a deflationary decade, and this week compressed about a year of it into seven days. On Monday a single OpenAI model released more mathematical research in one GitHub commit than most departments produce in a career. Days later OpenAI priced near frontier intelligence at one fifth of its flagship rate. And Samsung researchers kept shrinking the models themselves, squeezing capable language models into less than one bit per weight. Three different teams, three different layers, one direction. Thinking is getting cheaper to produce, cheaper to buy, and cheaper to carry in your pocket.
Why does this story matter?
Put the three stories together and the pattern glows. A model that writes 722 math papers in a week, frontier intelligence at a fifth of the price, and serious models running in under a gigabyte. The cost of a good thought keeps collapsing, and every collapse moves power from the owners of giant data centers to the builders with a laptop and a sharp idea. For readers that means the same thing it has meant every cycle, learn the tools while they are cheap, build while the giants are busy arguing about supply, and remember the advantage always lived in the clarity of the mind using it.
Sources
- AIWeekly on the 722 OpenAI math preprints
- Tech Insider on the withdrawn math papers
- Gadgets360 on GPT 6.1 Sol launch and pricing
- Forkast on the one fifth pricing move
New to crypto? Read the crypto glossary, browse frequent questions, read our story, or explore the story archive.