Fernando Vidal
@fernavid
Followers
4K
Following
3K
Media
235
Statuses
875
Cofounder and Chief Data Scientist at @3F_Research. Mainly trying to post cool charts. PM $FCTE $RAA https://t.co/H7PivYC5Tx
San Francisco, CA
Joined December 2012
RT @WarrenPies: S&P 500 - SENTIMENT. Now touching extreme optimism. Would be logical to see the market cool between mid-August and mid-Octo….
0
132
0
Big moves in individual stocks are happening at a historically high rate. Chart shows the percent of Russell 3000 stocks that have moved up more than 20% in a single day over the past 252 trading days.
3
3
10
The new Grok4 model still fails at this.
Interesting random multimodal AI model limitation, no model I have tried so far can read the Fed dot plot chart properly and turn it into structured data (counts by level by year, say). I tried o3-pro, gemini 2.5, claude opus 4. Weird right?
0
0
6
RT @WarrenPies: AI adoption by U.S. businesses taking off. Based on @fernavid work, daily tokens generated to ~4X by the end of 2026. So m….
0
46
0
Interesting random multimodal AI model limitation, no model I have tried so far can read the Fed dot plot chart properly and turn it into structured data (counts by level by year, say). I tried o3-pro, gemini 2.5, claude opus 4. Weird right?
3
0
10
A quick demo of the chart catalog feature on the 3Fourteen Research client site. Along with archives of every report we have ever written, every chart we have ever featured in our research is searchable and available with a link back to the report it was originally featured in on
1
1
21
RT @TheCompoundNews: End of an Empire | TCAF 188. @Downtown and @michaelbatnick are joined by @WarrenPies of @3F_Research to discuss what e….
0
25
0
I've been seeing this chart on % of consumer spending from the top 10% by income reaching 50%. Anyone know how they came up with this number? BLS Consumer Expenditure Survey suggests it's far less, with the top 20% being less than 40% (and top 10% being around 23%).
2
1
10
Competent open source models are a real shot in the arm to GPU demand.
On-demand H100s have gone from being available 95%ish of the time in early January before the release of Deepseek R1 to only being available 15% of the time today per this great dataset from @3F_Research
3
4
38
Seeing so many posts about Washington DC home listings EXPLODING but there is no context for how it differs from any random day in the past. Zillow new listing data as of December shows nothing interesting happening in DC, maybe the January update will show something.
2
0
8
From a recent o3-mini chain of thought summary:. "I'm showcasing the last flourishing self-confidence moment before a motivational dip in self-esteem, revealing a more nuanced and mature approach to the concept.".
0
0
2
DeepSeek's R1 has made it WAY harder to rent a cheap Nvidia H100 GPU. Hasn't been this hard to get a cheap H100 since early 2024. Curious how much open-source model usage is cannibalizing OpenAI/Anthropic's utilization vs growing the pie though.
0
0
11
Our clients will find this interesting in the context of our chart of the week in this week's model update.
H100 Pricing 🚀 - Jevon's Paradox in Action!. H100 prices exploded after the introduction of V3, as GPUs are starting to be monetized at much higher rates. More intelligence for cheaper means more demand. This is a meaningful shift to sluggish H100 spot pricing previously $nvda
0
0
6
Curious to see if it becomes harder to rent a cluster now that anyone can grab one and run their own private copy of DeepSeek on it.
Clients check your emails to hear our thoughts on the latest AI developments and market implications. Also, we discuss why will be watching our GPU availability metrics in coming months (below).
1
1
16
Maybe it’s already trying this out in R1!.
td;dr: I've identified dozens of "glitch" tokens which, when mentioned to DeepSeek, cause strange and unpredictable behavior. They each behave in unique ways, ranging from being "unspeakable" to inducing endless repetition and confusion.
0
0
2
Reasoning patterns learned with RL (DeepSeek) are cool, but imagine if the model could invent its own language rather than be held back by having to think in English. Humans did this with mathematical notation to great effect, what more efficient notation can AI come up with?.
2
0
7
Adding DeepSeek R1 to our chart showing AI efficiency gains over time. Y axis ~= performance, X axis ~= cost to run. Each research step forward lets us do more with less. The future of GPU demand remains a question of whether adoption or efficiency increases faster.
3
5
50
Context for this stat.
The S&P 500 has now closed the day at new ALL TIME HIGHS 47 times so far in 2024.
1
0
13
RT @TheCompoundNews: Strength Leads to Strength | TCAF 162🎙️. @Downtown and @michaelbatnick are joined by @WarrenPies and @fernavid of @3F_….
0
13
0
RT @CNBCClosingBell: "[Long-term] rates are one of the big risks to markets" says @3F_Research's @WarrenPies. "I think maybe the market has….
0
13
0