
14 y/o founder of @ThuemmlerA947 in Germany and developer of Nachhilfe Mentor App
Berlin, Deutschland
Joined July 2023
- Tweets1.6K
- Following186
- Followers41
- Likes956
Grok Bot can now book your flights, hotels AND rental cars Everyone says they'd never let AI book a trip They're missing the point. Booking is the easy part, the real value is planning it all in one chat My go-to for work and life. Here's how it works:
2
2
36
setting up @DevinAi bots to rebuild our adaptive BKT learning paths for Lyn AI Tutor. hype to see what an always-on software engineer teammate can ship
devin.ai/bots
23
1.7 to 1.25 in one year (German grades, 1.0 is best). I changed one thing: a skills system that makes AI teach me my way. Comment "learn" and I'll DM you my exact process. Follow me first, or DMs won't reach you.
1
29
LangChain raised a $300M Series C at a $3B valuation, led by CapitalG.
2.4x the $1.25B round a year ago.
Now sells agent management (LangSmith), model and cloud agnostic. 7,000 customers incl. Harvey, Rippling, Nvidia, Toyota.
Open source framework turned ops layer.
1
1
54
Nvidia CEO Jensen Huang has resigned after 33 years. CFO Colette Kress is interim CEO.
A petition to reinstate him passed 115,000 signatures by Saturday: 5,965 current Nvidia employees, 63,063 shareholders. Musk, Garry Tan, Tim Sweeney, Amjad Masad signed.
1
123
Why AI startups are nervous: Nvidia joined 105 funding rounds this year alone (PitchBook), often as largest investor.
Backed CoreWeave, Crusoe, Mistral, Lambda, Black Forest Labs. $800M into Reflection AI's $2B Series B. Hugging Face deal: $13B, closing 2027.
1
84
VCs told TechCrunch most startup bets came from Jensen himself. One investor on Kress: "She's much more conservative."
Expect fewer early-stage checks for neoclouds, inference and open-weight labs, short term.
techcrunch.com/2026/11/13/je…
1
56
My review agent never sees the writer agent's notes or sources. Only the brief, the drafts and a yes/no checklist.
First run, it failed 2 of 3 drafts.
One turned X's ranking weights into "1 reply = 7 likes", the exact misreading that draft was meant to warn about.
1
15
Growth guides love this one: "A reply the author engages with is worth 150 likes."
That's 75 ÷ 0.5 from Twitter's 2023 weights, and it was a misread even then. Those weights scale the predicted chance of it happening, not counts.
X's current For You scorer has no such term.
1
1
25
Current weights: github.com/xai-org/x-algorit…, home-mixer/params/param.rs. Reply 5.0 (20 when viewer and author are mutuals), like 0.708, no author-engaged-reply term. The comments there warn against reading weight ratios as "count equivalences".
20
Anthropic published how it builds safety evals.
Core point: "safer than the last model" only means it scored better on an eval someone built. The claim is only as good as the eval.
They borrow psychology's replication fixes: preregistration, better measurement.
2
1
29
Refusal rate is often the wrong metric. A model that refuses every self-harm prompt scores perfectly and still fails people in crisis.
Anthropic reports an "appropriate response rate" instead: share of conversations handled right in every turn. One bad turn fails the convo.
1
24
Example: pre-release transcripts for Fable 5.1 showed Claude implicitly validating self-harm as coping. A system prompt line helped: 60% on the API vs 94% on claude.ai.
Evals also run sandboxed, no internet.
anthropic.com/research/build…
11
Chinese open-weight models were allegedly trained on export-banned Blackwell GPUs.
Will DePue (ex-OpenAI, Unemployment Labs) analyzed the weights of DeepSeek, Moonshot, MiniMax and Ant Group models. His claim: Kimi K3 and DeepSeek V4 carry Blackwell training fingerprints.
3
1
53
Caveat (TechCrunch): this doesn't prove a law was broken. The chips could sit in an offshore data center. Epoch AI counts ~275k Nvidia accelerators imported to Malaysia in 2024-25.
Controls cover where chips ship, not who trains on them.
techcrunch.com/2026/11/13/ex…
25
Amazon reportedly plans a "World Product Graph": a database of every product sold anywhere online, including products Amazon doesn't sell (Reuters via The Verge).
Goal: keep shoppers and AI shopping agents on Amazon instead of Reddit or Google for product research.
1
17
Why: an internal Amazon doc says "Many Rufus answers are factually wrong or incomplete", and running shopper searches through the AI is expensive.
More product data should make answers more accurate and let Amazon use cheaper models.
1
12
Caveats: it depends on figuring out how brands organize their own sites (likely crawling other retailers). The doc lists an Oct 1, 2027 release date. Amazon calls it an outdated brainstorming doc.
Source: theverge.com/tech/1033697/am…
5