The Tape — DeepSeek, a Monday, and a Walk I Will Not Freeze

January 20, 2025. DeepSeek released R1, in the rooms that walk a Friday paper. A reasoning model. A WeChat walk put it twenty to fifty times cheaper to use than OpenAI's o1, depending on the task. I will keep January 20. I will keep twenty to fifty as a walk. I will not freeze a multiple as a finding I sat in a tokenizer. A multiple is a mouth. A mouth is not a holding.

January 10, in the same week's weather: DeepSeek-V3. A paper walk put training under $5.6 million, on Nvidia H800s, in about two months. Bernstein's Stacy Rasgon, in the rooms that bother with a caveat, said $5.6 million does not include prior research, architectures, algorithms, or data. I will keep $5.6 million as a walk. I will keep does not include. I will not add $5.6 million to another book's training walk. Two papers. One January. Labels are honest when they stay labels.

Then Monday arrived. January 27, 2025. Nvidia closed down just under 17 percent, in the Reuters weather; about 16.9 percent in a Yahoo Finance walk of Friday-to-Monday. Reuters, citing LSEG: about $593 billion of market value evaporated, a record one-day loss for any company on Wall Street. Yahoo: $589 billion, largest single-day loss in stock-market history. TechCrunch and Fortune walked nearly $600 billion and roughly $600 billion. I will keep the three walks. I will not add them. I will not freeze a tape as a finding that a Friday paper caused a Monday. A tape is a public. A public is not a holding that I sat in a pit.

The Friday-to-Monday walk, in TechCrunch's rooms: Nvidia closed $142.62 on Friday and $118.58 on Monday. I will keep both as a walk of two prints. I will not add them to Reuters's $118.42. Three prints. One Monday. A grocery line will try to spend a print as if I had a Bloomberg. I held a tape.

Nvidia's Monday loss, Reuters said, was more than double the previous one-day record, set by Nvidia the prior September. I will keep more than double as a walk. I will not invent the September print as this chapter's lead. A prior record is a sideboard. A sideboard is not a Friday paper.

The year-to-date walk in that Reuters: down about 11.8 percent for 2025 so far, after rising about 171 percent in 2024 and about 239 percent in 2023, trading at about 56 times earnings. After-hours, up about 2.5 percent Monday. I will keep 56 times as a walk. I will keep after-hours as a walk. I will not freeze a multiple as a finding that a paper changed a multiple. A multiple is a tape's costume. A costume is not a holding.

DeepSeek-V3, launched January 10 in the Reuters weather, used Nvidia's lower-capability H800 chips for training, in a paper last month, in that telling. I will keep H800. I will keep lower-capability as a walk. I will not remake another book's export-control kitchen as this H800 if that kitchen exists in a later clerk. An H800 is a chip a paper named. A paper's chip is not a finding that I sat at a BIS. I will not typeset an export you can invert. A museum sells a ticket.

Liang Wenfeng, in the rooms that walk a founder, sat behind High-Flyer, a quantitative shop, in those walks. I will keep the name as a walk of a public. I will not invent a desk. Inventing a desk is a kit. A founder is a public. A public is not a tour of a fund.

Distillation, in the later-weather rooms: U.S. labs walked allegations that a cheaper model had learned from a dearer model's answers. I will keep later. I will keep allegations. I will not upgrade an allegation to a finding I sat in a log. An allegation is weather. Weather is not a caption.

GTC 2025, later that March, in the rooms that walk a keynote: Jensen Huang said last year almost the entire world got it wrong. The computation requirement, the scaling law, more resilient, hyperaccelerated. The amount of computation needed as a result of agentic AI and reasoning is easily a hundred times more than we thought we needed this time last year. I will keep hundred times. I will keep got it wrong as a mouth. I will not freeze a hundred as a finding that a Monday was a shrug. A hundred is a keynote. A keynote is not a tape. Two seasons. One chip. Labels are honest when they stay labels.

CNBC with Jim Cramer, in a Wednesday walk of that conference: Huang called R1 fantastic, the first open-sourced reasoning model; it breaks problems step-by-step, provides multiple answers, verifies conclusions; this reasoning AI consumes 100 times more compute than a non-reasoning AI — the opposite of what many had expected. I will keep fantastic. I will keep open-sourced. I will keep opposite. I will not upgrade an opposite to a finding that a 17 percent unsayed itself. An opposite is a mouth in March. A 17 is a close in January. Two months. I will not add them.

He predicted global computing capital expenditures would reach $1 trillion by the end of the decade, majority toward AI, in that Cramer walk. I will keep $1 trillion as a walk. I will not add a trillion to $593 billion. A decade is a mouth. A Monday is a tape. Two kitchens. I will not add them.

Blackwell Ultra and Vera Rubin, in the keynote weather, as chips a reasoning model would benefit from. I will keep the names as a keynote's names. I will not remake another book's Closed division as this CUDA. That kitchen had an MLPerf. This kitchen has a hundred times. Two kitchens. I will not add them.

Ivo wrote GTC 2025 / hundred times / got it wrong / Cramer fantastic open-sourced / $1T decade walk / Blackwell Vera Rubin named / not a shrug of Jan 27 and boxed not a shrug. He typed a flyer that had used the word reasoning. He said reasoning is a brochure until a tape writes a close. I said reasoning sat in a March next to a January.

After-hours Monday, up about 2.5 percent, I already kept. I am walking it again so a grocery line cannot spend a 2.5 as a 17 that unsayed itself. A 2.5 is an after. A 17 is a close. Two prints. One Monday. Labels are honest when they stay labels.

Other chip names fell that Monday, in the rooms that walk a rout. I will not invent which. Inventing a ticker is a kit. A rout is a Reuters. A Reuters is not a finding that I sat in a book.

The App Store, in walks of that Monday, said a free assistant had overtaken ChatGPT in downloads. I will keep overtaken as a walk. I am walking free so a grocery line cannot spend a download as a subscriber I sat. A free is a walk. A walk is not a Chegg.

o1, in the WeChat weather, was the dearer cousin the twenty-to-fifty pointed at. I will keep o1 as a walk's noun. I will not remake another book's o1 as a product kitchen. An o1 is a multiple's denominator. A denominator is not a demo.

Ivo wrote after-hours 2.5 not a 17 / rout unnamed tickers / free assistant / o1 denominator / still not CoreWeave / still not split and boxed denominator. He typed a flyer that had used the word breakthrough. He said breakthrough is a brochure until a tape writes a close. I said breakthrough sat in an Andreessen next to an amazing.

Nerys asked if an o1 was a stain. I said it was a denominator. She said denominators are for fractions. I said this denominator is a walk. She wiped the eleven.

Nerys asked if a hundred times was a stain that spread from a 17. I said they were two mouths. She wiped the eleven. I did not write 100x on the sleeve.

The Friday paper had a title the rooms dated: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. I will keep the title as a paper's title. I will not freeze a title as a finding that a Monday sat inside an abstract. A title is a public. A public is not a tape.

DeepSeek-R1-Zero, in that paper: trained by large-scale reinforcement learning without supervised fine-tuning as a preliminary step. I will keep R1-Zero. I will keep without supervised fine-tuning as a paper's without. I will not typeset a without you can run. A museum sells a ticket. A ticket is not a recipe.

The paper's framework name sat as GRPO. I will keep GRPO as a paper's noun. I will not invent a gradient I sat. Inventing a gradient is a kit.

R1-Zero's AIME 2024 pass@1, in the paper's table: from 15.6 percent at the start of that RL weather to 71.0 percent. Majority voting, in the same page: 86.7 percent, matching o1-0912 in that telling. I will keep 15.6 / 71.0 / 86.7 as a paper's three. I will not add them. Three scores. One Zero. A grocery line will try to spend a sum as if I had proctored a contest. I held a table.

R1 itself, in the paper's later table: 79.8 percent Pass@1 on AIME 2024, slightly surpassing o1-1217 in that telling; 97.3 percent on MATH-500, on par with o1-1217 in the same mouth; 2,029 Elo on Codeforces, outperforming 96.3 percent of human participants, in that walk. I will keep 79.8. I will keep 97.3. I will keep 2,029 as a paper's Elo. I will not freeze an Elo as a finding I sat in a handle. A paper can spend a rating. A paper is not a holding that I counted a person.

Six dense distill checkpoints, in the paper's release weather: 1.5B, 7B, 8B, 14B, 32B, 70B, based on Qwen2.5 and Llama3 series. Distill-Qwen-7B: 55.5 percent on AIME 2024, surpassing QwQ-32B-Preview in that table. Distill-Qwen-32B: 72.6 percent AIME, 94.3 percent MATH-500, 57.2 percent LiveCodeBench, comparable to o1-mini in the paper's mouth. I will keep the six sizes as a paper's six. I will keep 55.5 and 72.6 as walks inside a table. I will not add a 7B to a 32B. Two checkpoints. One Friday.

I will not remake another book's Llama 2 seven-hundred-million-download kitchen as this Llama-3.3 base. That kitchen had a download and a Closed. This kitchen has a distill checkpoint and a table. Two kitchens. I will not add them. I will not restage an MLPerf.

MIT License, in the GitHub rooms that walk the weights: the code repository and the model weights, commercial use allowed, modifications and derivative works including distillation, in that telling; Qwen-derived checkpoints still carrying an Apache 2.0 parent, Llama-derived checkpoints still carrying a Llama license, in the same footnote weather. I will keep MIT as a page's license. I will keep the parent footnotes as footnotes. I will not typeset a license you can paste as a kit. A museum sells a ticket.

Poor readability and language mixing, the paper said of R1-Zero. Cold-start data — thousands of long chain-of-thought examples — before the R1 pipeline, to make a readable pattern. I will keep language mixing. I will keep cold-start as a paper's noun. I will not invent a mix I sat. A mix is a paper's mix. A paper's mix is not a Monday.

The paper walked post-processing by human annotators on some of that cold-start. I will keep annotators as a paper's noun. I will not remake another book's rater kitchen as this readability. That kitchen had a Scale and a McKinney. This kitchen has a paper and a footnote. Two kitchens. I will not add them.

An "aha moment," the paper named, in a table of an intermediate Zero: the model reallocating thinking time, in an anthropomorphic tone the authors boxed as interesting. I will keep aha as a paper's caption. I will not upgrade a caption to a finding I sat in a lab. A caption is a table. A table is not a tour.

Ivo wrote title Incentivizing / R1-Zero no SFT / GRPO / AIME 15.6-71.0-86.7 / R1 79.8 97.3 2029 Elo / six distill / MIT + parent footnotes / not Llama-2 700M / not Scale and boxed paper not tape. He typed a flyer that had used the word incentivizing. He said incentivizing is a brochure until a close writes a Monday. I said incentivizing sat in a title next to a Zero.

arXiv 2501.12948, in the rooms that walk a number next to a Friday. I will keep the number as a paper's number. I will not freeze a number as a finding that a Monday sat in an identifier. An identifier is a clerk. A clerk is not a tape.

Comparable to o1-1217 on reasoning tasks, the paper said of R1. I will keep comparable as a paper's comparable. I am walking 1217 so a grocery line cannot spend a 79.8 as an o1 I sat. A 1217 is a paper's cousin. A cousin is not a demo I will remake.

Ivo wrote 2501.12948 / o1-1217 comparable / still not a Monday I sat and boxed identifier. He typed a flyer that had used the word comparable. He said comparable is a brochure until a table writes a 79.8. I said comparable sat in a 1217 next to a Friday.

Thinking time, the paper said, rose through the RL weather: Zero learned to spend more tokens on a problem — hundreds to thousands of reasoning tokens — without an external instruction to linger. Reflection and alternative approaches emerged, in that telling, as test-time compute increased. I will keep hundreds to thousands as a paper's range. I will keep emerged. I will not typeset a linger you can prompt. A museum sells a ticket. A ticket is not a chain.

Ivo wrote thinking time rose / hundreds-thousands tokens / reflect emerge / still not a recipe and boxed not a recipe. He typed a flyer that had used the word linger. He said linger is a brochure until a table writes a 71. I said linger sat in a Zero next to a token.

Nerys asked if a token was a stain. I said it was a paper's range. She wiped the eleven. I did not write token on the sleeve.

Nerys asked if a 1217 was a stain. I said it was a cousin. She wiped the eleven.

Nerys asked if a Zero was a stain that spread from a V3. I said they were two papers in one January. She wiped the eleven. I did not write Zero on the sleeve.

The V3 architecture sitting under R1, in the rooms that walk a mixture: 671 billion parameters, sparse, about 37 billion active per forward pass. I will keep 671 billion as a walk of a V3 that a Friday paper sat on. I will keep 37 billion active as a walk. I will not add them. A total is a costume. An active is a costume. Two costumes. One Base. A grocery line will try to spend a 671 as if I had counted a weight. I held a walk.

Launch API pricing, in later rooms of that January: $0.55 per million input tokens on a cache miss, $0.14 on a cache hit, $2.19 per million output, deepseek-reasoner as the endpoint name. I will keep $0.55 / $2.19 as a launch walk. I will keep $0.14 as a hit walk. I will not add them to $5.6 million. A price is a menu. A $5.6 million is a V3 training walk Bernstein already caveated. Two menus. One January. I will not typeset a call you can run. A museum sells a ticket.

The WeChat twenty-to-fifty sat next to o1 as a multiple. I already kept the multiple. I am walking the menu so a grocery line cannot spend a twenty as a $0.55 I sat. A twenty is a mouth. A $0.55 is a later clerk of a menu. Two walks. One cheaper. Labels are honest when they stay labels.

128,000-token context, in those later spec walks. I will keep 128,000 as a walk. I will not invent a window I sat. Inventing a window is a kit.

Ivo wrote 671B / 37B active V3-under-R1 / $0.55 $2.19 launch / $0.14 hit / 128k ctx / do not add to $5.6M / still not a tape and boxed menu not paper. He typed a flyer that had used the word reasoner. He said reasoner is a brochure until a close writes a Monday. I said reasoner sat in a menu next to a miss.

Nerys asked if a 671 was a stain that spread from a 5.6. I said they were two walks in two costumes. She wiped the eleven. I did not write 671 on the sleeve.

Hugging Face hosted the weights, in the rooms that walk a download page next to an MIT. I will keep Hugging Face as a walk of a host. I will not freeze a host as a census I counted. A host is a page. A page is not a Monday.

The first open research to validate that reasoning can be incentivized purely through RL, without SFT, the paper said of Zero. I already kept without SFT. I am walking first open so a grocery line cannot spend a first as a finding I sat in a lab. A first is a paper's first. A paper's first is not a tape.

Ivo wrote HF host / first open RL-no-SFT mouth / still not a Monday and boxed host. He typed a flyer that had used the word validate. He said validate is a brochure until a close writes a 17. I said validate sat in a Zero next to a Face.

Nerys asked if a Face was a stain. I said it was a host. She wiped the eleven.

Vaughn asked if I was scoring an Elo. I said I was dating a table. He dated the pad. Tables are honest. Elos are a later costume I will not wear as a kit.

Quill asked whether "got it wrong" meant "they peaked." I said peaked is a polo word. He wrote a five.

Bram said correction hundred adjacency. I said a hundred is a keynote, not a vest. He zipped a pocket.

Ivo wrote Fri 142.62 / Mon 118.58 walk / more than double prior Sept record / 56x walk / H800 walk / not BIS kit / distillation later allegation and boxed allegation. He typed a flyer that had used the word efficient. He said efficient is a brochure until a tape writes a close. I said efficient sat in a WeChat next to a twenty-to-fifty.

800k samples, the paper said, curated with R1, used to fine-tune the smaller dense models. I will keep 800k as a paper's 800k. I will not invent a sample you can paste. Inventing a sample is a kit.

Self-verification, reflection, and long chains of thought, the paper walked as behaviors that emerged in Zero. I will keep the three as a paper's three. I will not typeset a reflection you can prompt. A museum sells a ticket.

Rule-based rewards — outcome and format — and a refusal of a neural reward model because of reward hacking, in that methods weather. I will keep rule-based. I will keep reward hacking as a paper's fear. I will not invent a reward I sat. A fear is a methods page. A methods page is not a finding I held in a pit.

DeepSeek-V3-Base as the starting checkpoint for both Zero and R1, in the paper's pipeline. I already kept V3 as a January 10 training walk. I am walking Base so a grocery line cannot spend a $5.6 million as this RL. A Base is a paper's Base. A $5.6 million is a V3 walk Bernstein already caveated. Two rooms. One January. I will not add them.

Ivo wrote 800k distill samples / self-verify reflect long CoT / rule-based reward / V3-Base start / still not a tape I sat and boxed Base. He typed a flyer that had used the word emerged. He said emerged is a brochure until a table writes a 71. I said emerged sat in a Zero next to a mixing.

Nerys asked if an 800k was a stain. I said it was a paper's count. She said counts are for sleeves. I said this count is a sample I will not invent as a person. She wiped the eleven.

Nerys asked if a double was a stain that spread from a 17. I said they were two walks. She said walks are for dogs. I said these walks are a record I did not count. She wiped the eleven. I did not write 593 on the sleeve. Writing 593 on a sleeve is a kit.

The Nasdaq fell about 3.1 percent that Monday, in the Reuters walk. DeepSeek's free assistant had overtaken ChatGPT in downloads from Apple's App Store, in that telling. I already kept overtaken. I will not freeze an App Store as a census I counted. A chart is a walk. A walk is not a finding that I sat in Cupertino.

Nvidia, in a statement walk: R1 was an excellent AI advancement. Marc Andreessen, in a mouth the rooms repeated: one of the most amazing and impressive breakthroughs I've ever seen. I will keep excellent. I will keep amazing as a mouth. I will not upgrade a mouth to a finding that a paper is a product I sat. A mouth can praise a Friday and still not be a Monday.

Reuters walked a close of $118.42. CNBC walked $118.58. I will keep both as walks. I will not add them into a novel that says I held a print. Two closes. One Monday. A grocery line will try to spend a close as if I had a terminal. I held a tape.

I printed the closes. Nerys asked if 17 was a score. I said it was a walk. She said walks are for dogs. I said this walk is a chip. She wiped around the eleven.

Ivo wrote R1 Jan 20 / V3 Jan 10 $5.6M walk / Bernstein does not include / Jan 27 ~17% / $589 / $593 / ~$600 do not add / App Store overtaken walk / not CoreWeave $2B / not DOJ split and boxed do not add. He typed a flyer that had used the word correction. He said correction is a brochure until a tape writes a close. I said correction sat in a vest last Tuesday.

I will not remake another book's CoreWeave line as this Monday. That kitchen had a residual order and a 67. This kitchen has a Friday paper and a tape. Two kitchens. I will not add them.

I will not remake another book's chip inquiry as this close. That kitchen had a June 6 split and an 80 percent walk. This kitchen has a January 27 and a 17. Two kitchens. I will not add them.

Paz asked if I was oiling a chip. I said I was dating a tape. He dated the bottle. Tapes are honest. Oils are a sacrament.

Quill asked whether "seventeen" meant "they peaked." I said peaked is a polo word. He wrote a five.

Bram said correction tape adjacency. I said a tape is a Monday, not a vest. He zipped a pocket.

Nils asked if I was filing the 593. I said I was dating a walk. He dated the stack. Walks are honest. Filings are a later costume.

The next room is a lease an analyst said was canceled. The seventeen is still this one.

I will not give you a training run you can copy. I will not give you a prompt that summons an R1. I will tell you January 10 happened as a V3 walk, that January 20 happened as an R1, that January 27 happened as a tape, that three market-cap walks sat next to one Monday and I will not add them, that an App Store walk said overtaken, and that a shop in a rented suite still pays an invoice after a feature paused because confidence is a typeface. The next room is a megawatt a note named. The tape is still this one.