DEEPSEEK SCOOP
V3.1 Drops is it Copycat AI?
You might have seen me using Manus and selling a horse in Texas using China’s DeepSeek. (I filmed it.)
Since April, they have grown.
THIS WEEK they dropped DeepSeek V3.1, a 685 BILLION PARAMETER model.
Are these just overseas clones on GPT [like fake Nike but bigger]? OAI has basically said as much.
In the trend of “quiet launches” (sorta like I did with Jurisphere) DeepSeek quietly unveiled DeepSeek-V3.1, an upgraded version of its flagship large language model (LLM). The new boasts enhanced capabilities, including a expanded 128,000-token context window.
+ + + 128,000-token context window refers to the maximum amount of text a large language model (LLM) like DeepSeek-V3.1 can process and "remember" in a single interaction.
= Tokens are the basic units of text—roughly equivalent to subwords (the example I keep giving is the three-letter word “the”) — "unbelievable" might split into "un," "believ," and "able").
= 400 pages of input/output at once, 96,000 words.
= enables complex tasks like analyzing long documents, maintaining extended conversations, or solving multi-step problems without losing context. A key advancement for efficiency in AI applications.
BUT CYRUS, BUT CYRUS, DIDNT YOU SHOW US LONGWRITER EXACTLY A YEAR AGO?
V3.1 also introduces a “hybrid inference structure”, allowing for flexible precision adjustments, and has been optimized for domestic Chinese hardware, signaling a push toward semiconductor independence amid global chip tensions.
If one is into benchmarking, and to me these are mostly paid, dubious, or both, DeepSeek touts that V3.1 it scores 71.6% on the Aider benchmark for coding tasks and shows marked improvements in math, reasoning, and agentic capabilities over its predecessor, V3.
Released open-source on Hugging Face with minimal fanfare—no splashy announcements or detailed whitepapers—the model is now accessible via DeepSeek's web platform, app, mini-program, and API, maintaining backward compatibility for seamless integration.
Some say and will say it is the most powerful open AI model yet, challenging Western giants like OpenAI and Anthropic.
DeepSeek had been expected to release its “R2” model this year but didn’t. Speculation is that there was dispute between it’s intended provider for chips (“Ascend 910 B” by Huawei) and thus DS pivoted back to “AI’s home base” — Nvidia hardware.
Future whispers suggest an upcoming "omnimodal" V4 or even a full R2 drop, potentially bundled with a technical paper, as DeepSeek conservative approach—testing features internally before public release—hints at bigger plays ahead.
I am sorta cracking up xAI launched Grok 4 / recently 4.20 / 5 is reportedly coming this year………………………………..looks like copycatting to me.
