Luis Herrera (@lherrera) header
Luis Herrera (@lherrera) profile picture

No plan survives first contact with the reality. Opinions are my own and not the views of my employer. RTs ≠ endorsements

Madrid
Joined March 2011
Photo by Luis Herrera (@lherrera)Photo by Luis Herrera (@lherrera)Photo by Luis Herrera (@lherrera)Photo by Luis Herrera (@lherrera)Photo by Luis Herrera (@lherrera)Photo by Luis Herrera (@lherrera)Photo by Luis Herrera (@lherrera)Photo by Luis Herrera (@lherrera)Photo by Luis Herrera (@lherrera)Photo by Luis Herrera (@lherrera)
Luis Herrera retweeted
Snowflake has four levels to it: Level 1: Snowflake Fundamentals - Understanding Snowflake SQL - Creating databases, schemas, and tables - Loading data with COPY INTO - Understanding virtual warehouses - Basic query optimization Level 2: Analytics Engineering - Building data pipelines with Snowpark - Using dbt for transformations and testing - Implementing incremental models and dynamic tables - Understanding micro-partitions and clustering - Optimizing compute costs and query performance Level 3: Advanced Data Engineering - Building streaming pipelines with Snowpipe Streaming - Implementing CDC and near-real-time ingestion - Working with Iceberg tables and open table formats - Orchestrating complex data workflows - Designing scalable, reliable data architectures Level 4: AI-Native Data Engineering - Building AI applications with Cortex AI - Processing unstructured data with LLMs - Implementing vector search and RAG pipelines - Building autonomous agents with Cortex Agents - Combining structured and unstructured data for AI-powered analytics Most Snowflake engineers are stuck at Level 1 or 2. But the platform is rapidly expanding into Level 3 and 4. In 2027, knowing SQL won't be enough. The engineers who understand how to build end-to-end data and AI systems on Snowflake will have a massive advantage. Snowflake isn't just a data warehouse anymore. It's has become a full-stack data and AI platform! What level are you at? I'm hosting a free lightning lesson about everything you need to learn on Snowflake tomorrow at 10 AM Pacific here: dataexpert.io/lightning-less…
9
22
1
158
8,321
Luis Herrera retweeted
Analytics engineering is catching the world on fire! Snowflake is now agentic with Coco dbt has a semantic layer! Data foundations still matter a lot. Knowing your Airflow DAG is producing correct data matters. Knowing that your data model and AI agent is doing the right actions also matters! We're launching a November 16th boot camp you can join here: dataexpert.io/program/snowfl…
9
10
1
121
7,274
Luis Herrera retweeted
Live at the Silicon Valley AI Hub tonight, marking one year of builders shipping together. #SVAIHub
5
2
17
2,183
Luis Herrera retweeted
To be clear, as we say in our latest Risk Report (teapawt.net/AnthropicAI/status/208…), I think the risk from present models is low. What I am worried about is superintelligence arising from recursive self-improvement, as we have said is happening faster than we thought (teapawt.net/AnthropicAI/status/206…).
As part of our Responsible Scaling Policy, we publish regular Risk Reports. These share detailed information on the risks of our systems and how prepared we are to address them. Our second Risk Report is now available: anthropic.com/aug-2026-risk-…
335
645
174
6,732
3,079,219
Luis Herrera retweeted
Jacob is correct here—we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade. I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.
Replying to @hilbertspaess
The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt. If anything, many executives and senior researchers will couch their phrasing in the press to sound sensible - but I hear the same people express fear privately. No other human activity poses this level of danger.
4,795
9,842
10,494
59,043
43,332,117
Luis Herrera retweeted
More model choice is coming to Snowflake Cortex AI with @OpenAI GPT-6 Astra. Now in private preview, GPT-6 Astra brings OpenAI’s latest reasoning and computer-use capabilities to Snowflake. Available across Snowflake CoCo, Snowflake CoWork, Cortex Agents, Cortex AI Functions and Cortex Inference. Dive in 👉 bit.ly/4xgr296
1
5
26
2,402
Luis Herrera retweeted
Across 253 CoCo sessions, 23-35% of agent turns were rediscovering what the team already knew. One pattern recurred in 24 separate sessions. In another study, Experience Cards distill review-rich merged PRs into seven-field local memory, retrieved at review time. Recall +4.6 to 7.8pp, cents per PR. The full writeup 👉 bit.ly/46A9zxh
2
4
23
2,355
Luis Herrera retweeted
Most AI agents are stuck in advisory mode; they can't execute. But the tide is turning: 44.5% of enterprises are investing in unified semantics, 35% in zero-copy storage. They're building an Agentic Control Plane, not just chatbots. Along with @TheFuturumGroup, we mapped the 4 pillars to bridge the gap. bit.ly/4yajMfz
3
5
18
2,259
Luis Herrera retweeted
The biggest model isn't always the best one for the job. Snowflake VP of AI Baris Gultekin on what it really costs to run AI at scale, why data strategy comes first, and where the tech still falls short. 👇
What does it actually cost to run AI? And why does the data behind it matter so much? Co-hosts @LeahBStern and @theofficial_ksa had a really interesting conversation with Baris Gultekin, Vice President of AI at @Snowflake , who has been working on AI for nearly two decades,
4
2
4
18
2,508
Luis Herrera retweeted
Semantic view materializations: one ALTER SEMANTIC VIEW ... ADD MATERIALIZATION and the query planner rewrites your queries to scan a precomputed rollup instead of the base tables. TPC-DS benchmark: 59–91x speedup. Data scanned: 359GB → 120MB. One materialization served all four queries. Full breakdown 👉 bit.ly/4ghwTWu
1
16
2,302
Luis Herrera retweeted
Migration work isn’t just code conversion. It’s dependencies, deployment order, data movement, validation, and a lot of manual reconciliation. Snowflake Data Superhero Abhishek Mittal breaks down how Snowflake AIM structures that workflow through Snowflake CoCo: bit.ly/3SMmQzB
1
2
1
15
2,474
Luis Herrera retweeted
Ladies and gentleman, CEO of Snowflake on GLM 5.2. If your CEO does not understand what Sridhar has written your company is not going to make it.
Early results from the @snowflake's coco team on GLM-5.2 vs Opus-4.7 on dbt-bench — what the trajectories actually show 🧵
7
36
2
954
341,271
Luis Herrera retweeted
Cursor just announced a new 1.5+ trillion parameter model pre-trained on over 100k GPUs. According to Cursor CEO it is: "as big as Opus and GPT" Yes, the cat is out of the bag and I can spoil the party. Opus 4.5 to 4.8 and GPT-5 to GPT-5.5 are not that big! (they are all below 2T params) This means that the current performance of GPT-5.5 and Opus 4.8 is achievable for open-source, because guess what else is that size? -> DeepSeek-V4-Pro The only moat is scaling. So far, Anthropic is the only lab that successfully made the jump to the ~10T scale, which is also why I no longer see OpenAI catching up until year end. Anthropic can just keep throwing RL compute at Mythos for the next 1-2 years and it will keep improving. Google went not quite as big and sparsity-maxxed a bit too much. They also have apparently no clue how to RL to make the model actually usable. OpenAI is still scarred from GPT-4.5 xAI and Meta are still planning.
And here’s the full video of @mntruell announcing Cursor’s new model at Compile.
123
160
32
2,946
493,954
Luis Herrera retweeted
at this pace, @cursor_ai will be the number one or two coding agent on a year very impressive squad... now with unlimited compute behind them the best acquisition since Instagram and YouTube
183
196
31
5,396
211,086
Luis Herrera retweeted
Working with agents should feel like working with a colleague. You should be able “speak to” them not just with text chats, but by gesturing at a screen together, talking live, etc.
With Design Mode, you can now point, draw, or talk to update your UI.
232
392
27
2,974
1,026,158
Luis Herrera retweeted
Lots to do together. Excited to be joining forces with @SpaceX to build useful AI.
SpaceX has exercised the option to acquire @cursor_ai in an all-stock transaction with the goal of building the world’s most useful AI models. For the past few months, SpaceXAI has been jointly training a model with Cursor, which will be released in Cursor and Grok Build soon. We look forward to working closely with the Cursor team to advance our frontier AI capabilities
719
881
130
13,223
3,984,230
Luis Herrera retweeted
SpaceX has exercised the option to acquire @cursor_ai in an all-stock transaction with the goal of building the world’s most useful AI models. For the past few months, SpaceXAI has been jointly training a model with Cursor, which will be released in Cursor and Grok Build soon. We look forward to working closely with the Cursor team to advance our frontier AI capabilities
SpaceXAI and @cursor_ai are now working closely together to create the world’s best coding and knowledge work AI. The combination of Cursor’s leading product and distribution to expert software engineers with SpaceX’s million H100 equivalent Colossus training supercomputer will allow us to build the world’s most useful models. Cursor has also given SpaceX the right to acquire Cursor later this year for $60 billion or pay $10 billion for our work together.
1,648
4,062
1,631
35,927
26,139,382
Luis Herrera retweeted
Claude + Reddit intelligence + LinkedIn data = Content OS that replaced a $14,500/month team. (48K followers. 500+ booked calls. $387K in attributed revenue) This 6-skill system eliminates writers, strategists, researchers, and schedulers using brand memory + real audience intelligence + calibrated AI writers... → No more $14,500/month payroll for a 4-person content team → No more 20+ hours weekly building content calendars from scratch → No more guessing which hooks actually stop the scroll → No more AI slop that sounds nothing like you → No more starting every post from a blank page Just 1 topic in → a finished month of content out. Here's how it works: → Brand Memory (your voice, proof, and exact buyer in one file. every skill runs off it. nothing generic ships.) → Reddit Intelligence (mines your ICP's subreddits for the exact phrases they use. posts start from real words, not blank pages.) → LinkedIn Viral Signal (hooks already stopping the scroll in your niche, last 30 days, ranked by real engagement.) → Four Calibrated Writers (thought leadership, lead magnet, story, value. one writer per stage of the buyer journey.) → Voice Firewall (7-dimension QA that scores and rewrites every draft before it ships under your name.) → Finished Board (every post handed over with its visual and first comment. ready to approve.) Built on real content infrastructure. Zero payroll. Consistent output. Enterprise-grade quality. Results from the system: - $14,500/month content team eliminated - 48K LinkedIn followers - 500+ calls booked from content - $387K in attributed revenue Want the complete 6-skill build? Like + comment "OS" + repost, and I'll DM it to you. (must be following)
532
220
4
735
63,802
Replying to @HedgieMarkets
The Gingerballmaxxing Paradox strikes again! Take that, Jevon. teapawt.net/Dr_Gingerballs/status/…
The way to think about AI has always been cost/benefit. If I take the sum total impact of what AI has solved to date, and divide it by the cost, does that number look amazing or disgusting? Stated another way, do the aggregate revenues exceed the costs, and will they ever? Stated yet another way: will productivity go up or down? The Erdos solution going around today is very cool. If it feels like the model found a needle in a haystack, that's because it did. I don't say that to dismiss the accomplishment, but to put it in context. And to do that, you have to understand how the solver works. The strategy itself is actually not new, or even particularly innovative. At the heart is the same search type algorithms that have fueled advances in GO and Chess solvers. You have a wide parameter space, an objective function, and constraints. In chess, the parameter space is all of the moves that can happen in any game possible. The constraints are the size of the board and the rules. The objective is to checkmate. With enough compute, you can "play" every possible game in your head and choose the most productive moves. Moving the knight here leads to 80% winning outcomes vs 60% if I do not move the knight. This is similar to counting cards. Even similar to IBM Watson playing Jeopardy! The strategy is largely the same. What has changed is the amount of money spent on compute to solve problems. In the recent Erdos math solution, the solver is also simple. As far as I can see, the field of math is about creating lemmas (proven steps) and then combining existing lemmas to arrive at larger conclusions. Creating lemmas can be difficult, as it requires abstraction of some observable feature of the universe. Stringing lemmas together to come to a conclusion is also difficult, but in a different way. It's less about abstraction and more about search. It's a game of chess. So you train a model on every lemma that exists, as well as which lemmas can be strung together. Much like defining puzzle pieces to a puzzle with a lot of different possible valid solutions. You then define the objective (how many paths of equal length can I draw between points on a grid). Then you search through all of the possible combinations of lemmas that can optimize that number. So the most crucial aspect of this, and what appears to be somewhat hidden, is the cost of the solution. If it was $1, that is mind blowingly revolutionary. If it was $1M, that is starting to get pricy. If it was $10M+, that is a pretty inefficient use of compute. If all scientific breakthroughs are worth it no matter the computational cost, then we should stop running LLMs and start running density functional theory (DFT) calculations, which determine atomic interactions from quantum mechanics. We have barely begun to even begin to fathom how much compute we would need to brute force all possible large ensemble atomic interactions over relevant time scales. But the outcomes are also potentially revolutionary: finding better materials, better drugs, better, well, everything. The benefits and the costs are infinity, and dividing the two is pretty meaningless. And herein lies the problem. The models aren't getting any more efficient, they are just getting bigger. And they cannot continue to get bigger, and more expensive, forever if we are going economically solve large problems with brute force. The only way this strategy works is if the cost of the hardware and the electricity comes down by many orders of magnitude. Renting the compute of a 1 GW datacenter for $1000 per day would be truly revolutionary for scientific discovery. Just the electricity to run that datacenter would be about $4M. The chips themselves would cost another $1-10M. So the cost of compute and energy needs to drop 5,000-15,000x. That's like buying a GB100 for $5. So the hardware and energy costs are going in the wrong direction currently for any of this to make sense. The argument that brute force search is somehow going to get cheaper in software is the big lie that AI labs are pushing to lure investors into buying more compute. The models have always just been different flavors of brute force, and the bottleneck is the cost of hardware. I have seen two trends on X lately. The first is ballmaxxing, where people inject saline into their testicles to make the pouch appear larger. The second is the discussion of paradoxes (Yes, Jevon, you have a weird name and we are sick of Satya screaming it from his goon cave). Therefore, I propose The Gingerballmaxxing Paradox (GBP): where the cost of compute proportionally scales with desire to brute force solutions, while the long term success of the brute force strategy requires the cost of compute to drop. The logical conclusion of the GBP is that productivity will continue to decline as we pursue brute force into increasing hardware prices, and we cannot see the desired compute renaissance until the entire AI investment crashes and is liquidated for next to nothing. I, for one, am looking forward to the crash, and intend to try and capture compute actually cheap enough to brute force some valuable things.
1
21
4,256