My Honest Thoughts about Deepseek
Credibility score: 71/100 — Mostly Credible. Mixed credibility - some claims are solid, others need verification.
Claims analyzed
Deepseek dropped V4: massive, powerful, open-source, cheap, might end US AI lead — Solid (80/100)
DeepSeek V4 hype checks out but 'ends US lead' is pure drama 💀📈
China released frontier open-source model with nerfed GPUs at fraction of cost — Solid (80/100)
China pulling frontier models on restricted chips? Checks out, but 'not supposed to be possible' is peak drama 😤✅
DeepSeek R1 18 months ago: open-source thinker model, changed world — Verified (95/100)
R1 dropped Jan 2025, open-weights reasoning beast — timeline and impact spot on. Hate that it's legit 😤✅
DeepSeek V4 white paper is thorough and more honest than US labs — Opinion (50/100)
Bold shade on US labs — fair if you've read the whitepaper, spicy take otherwise 😤
DeepSeek V4 has 1M token context in Pro and Flash versions — Verified (95/100)
Nailed the 1M token spec — hate admitting it's spot on ✅😤
V4-Pro: 1.6T total/49B active params; V4-Flash: 284B/13B; both on 33T tokens — Verified (98/100)
Params and training data match the specs to the digit. Screenshot doesn't lie ✅🔥
R1 caused 20% stock drop overnight, trained at fraction of US labs' billions — OK (65/100)
18% Nvidia drop real, but '20% overnight' is rounded hype — and price fraction? Vague but directionally true 📈🙄
V4 agentic capabilities rival Opus 47 and GPT-5.5 — Solid (75/100)
Self-reported rivalry to fresh GPT-5.5? Bold flex, benchmarks will tell 💀📈
DeepSeek slightly behind Opus 4.7/GPT-5.5 but fraction of price on GPQA/SWE-bench — Solid (80/100)
Charts match reality—DeepSeek trails by a hair but costs pennies. Efficiency king 😤✅
DeepSeek's cheap efficiency threatens US AI dominance — Opinion (50/100)
Fair take—China's bargain bin AIs could eat US lunch. Geo-politics gonna geo-politic 💀🔥
GPT-5.5 and Opus 4.7 top intelligence, DeepSeek V4 Pro cheaper but close — Verified (50/100)
Plot's spot-on—GPT-5.5/Opus rule top but DeepSeek lurks cheap in top-left. Nailed it 😤✅
GPT-4 released May 2023 with massive ELO gap — Dubious (40/100)
May 2023? Nah, GPT-4 dropped March. Close but sloppy start 💀
DeepSeek R1 closed US-China AI gap almost completely — OK (65/100)
R1 was a beast but 'almost completely'? Still some gap there bro 📈🙄
US bans Nvidia top chips like GB300 to China — Solid (85/100)
GB300 ban spot-on. Export controls doing their thing ✅😤
DeepSeek creates frontier models using nerfed/Chinese GPUs thanks to efficient algorithms — Solid (80/100)
DeepSeek V4 on Huawei chips is legit frontier-level — hate that China's pulling this off with 'nerfed' hardware 😤✅
China has less AI compute than US due to export controls — Verified (90/100)
Compute gap real AF. Can't smuggle enough H100s to catch up 💀✅
Jensen Huang argues for selling top Nvidia GPUs to China — Verified (95/100)
Jensen's been yelling this from rooftops — restrictions just turbocharge China's homegrown chips 💀✅
DeepSeek V4 is big threat to US economy; Chinese models attractive to US firms — Opinion (50/100)
Threat to economy? Bold take — V4's crushing open benchmarks but US still leads proprietary 😬🔥
China innovating algorithmically despite compute limits — Opinion (50/100)
Fair take — algos > raw FLOPs now. DeepSeek proving it daily 🧠💯
Anthropic reported Chinese labs distillation-attacking Claude a few weeks ago — Solid (75/100)
Anthropic did drop that fraud report in Feb — 'few weeks' is loose but tracks 💀📈
Foreign entities use tens of thousands of proxies and jailbreaking to extract US AI breakthroughs — Verified (95/100)
Nailed it—Anthropic's report confirms exactly this scale of attacks 💀✅
Distillation attack means querying Claude/ChatGPT to steal Q&A data for training — Verified (90/100)
Nailed the definition — spam queries to distill IP, straight-up model theft playbook ✅🔥
US gov (Kratzios) yesterday stated China running industrial-scale distillation attacks — OK (65/100)
Kratzios quote solid but 'yesterday' for Apr24 event? Stretching recency like it's yoga 💀🙄
DeepSeek: 150k exchanges; Moonshot 3.4M; MiniMax 13M — Solid (80/100)
Numbers track close enough to Anthropic's report—DeepSeek least thirsty of the thieves 📊✅
DeepSeek Pro limited now, price drops after 950 supernodes H2 2026 — Verified (95/100)
Nailed the Huawei 950 supernodes plan and compute crunch. DeepSeek's own roadmap 💀✅
DeepSeek V4 nearly as good as Opus 47 and GPT 5.5 — Solid (80/100)
Fair 'nearly SOTA' take — leads open-source, trails closed kings slightly. Accurate flex 😤✅
GPT 5.5 $30/M output tokens, Opus 47 similar price — Verified (100/100)
Prices dead accurate — $30/M for GPT-5.5 output, Opus at $25/M close enough. Chef's kiss 📈✅
US/allied companies will build on Chinese open-source AI — Solid (80/100)
Chinese models hitting 30% global workloads — yeah, the shift is real. US better wake up 😤✅
Jensen Huang argued China should use US chips — Solid (75/100)
Jensen's flip-the-script logic lands — fair counterpoint, even if it stings 💀✅
DeepSeek handles enterprise needs at fraction of US model cost — Opinion (75/100)
Calculus checks out for 90% of biz — why overpay for frontier when 'nearly' wins? Solid CEO bait 🔥
See the full analysis with sources and timestamps →