Creator perspective · MoneyYijia Liang explains that using ChatGPT or Claude off the shelf to process a 30 megabyte Excel model would be expensive and slow because it prints out every single piece of context about the model in code for the model to read (e.g., via OpenPyXL). Because Compound developed its own engine for Excel, he claims they are more accurate, faster, and cheaper, since the engine processes the model and lays it out in an easy-to-read format for the model, exposing a much richer set of APIs and tools for the agent to work over.
AI Engineer · Why We Rebuilt Microsoft Office From Scratch for AI Agents - Yijia Liang, Compound AI
Report · SocialKyle presents a conversation technique he claims to have created himself, in which short verbal acknowledgements of the previous sentence make a woman feel heard and get her to invest more in the conversation.
hedged
Coach Kyle · How to Make Women Feel Heard (and get them to open up)
Principle · MoneyWhen a production trace reveals a problem, the team logs it into the evaluation framework and asks the agent to continue auto-research on it as a single-question task; the agent identifies the problem, replicates the source trace, runs variants to verify the fix, adds a hill climb target describing the fix, and scores candidates against the production variant.
AI Engineer · How We Built an Agent That Improves Itself — Zubin Aysola, Weights & Biases
Creator perspective · MindIn clinical care, when the codependent person in a family lets go, surrenders, starts attending Al-Anon or Codependency Anonymous meetings, and takes care of themselves, the addicted loved one often gets into recovery or starts making moves toward it. It works like a tug-of-war rope: when one person drops their end, the other person stops pulling and may go get into recovery. Dr. Lembke describes this pattern as happening 'like clockwork' in her clinical experience.
Andrew Huberman · Overcome Addictions and Obsessions of All Kinds | Dr. Anna Lembke
Creator perspective · MoneyOne use case for fine-tuning is proprietary data sets that agents need to reason about, understand, and take action on, which are not present in any pre-training set. Out-of-the-box Claude and GPT would not work at all for these tasks, so teams had no alternative but to fine-tune open-source models.
AI Engineer · Training Your Own Agent Models With RL Is Now Practical - Jason Mancuso, Modal
Creator perspective · FuelA single cup of mushroom coffee almost certainly does not deliver enough lion's mane to produce cognitive results. If you need roughly at minimum 1,000 mg of lion's mane fruiting body daily to see an effect, most mushroom coffee blends probably do not get close to that.
probably
Also discussed by Dr. Ryan Young on Dr. Brian Yeung, ND (unclassified source — not counted until classified). Not counted as independent sources.
Victoria Dorsano · I Tried Ryze Mushroom Coffee for 30 Days (Honest Review)
Creator perspective · FuelRyze claims a 2,000 mg mushroom blend but does not disclose how much of each mushroom species is included, so consumers cannot know what dose of what mushroom makes up the total. Ryze also claims to use a mixture of mycelium and fruiting bodies without disclosing the proportions.
Consequential creator advice requires independent verification before Today can promote it.
Victoria Dorsano · I Tried Ryze Mushroom Coffee for 30 Days (Honest Review)
Creator perspective · MoneyDuvvuri explains reconciliation as a process where a provider's initially stated charge may not be true; e.g., OpenAI might say 100K for the month then come back and say it was 120K. The data is not constant and changes over time: providers change data that was previously promised not to change, so products must fetch new data that differs from the old. She endorses that you need a way to catch reconciliation adjustments and get one true, honest place to see all usage across all providers.
hedged
AI Engineer · Why No One Agrees on What Their AI Tokens Cost - Pratyusha Duvvuri, Cisco
Creator perspective · MoneyDuvvuri says it is extremely hard to bucket all ICPs into just three categories, but it is necessary to build a product. The three ICPs Cisco came up with are the engineer, the engineering manager, and the finance folks. Engineers worry about which models to use, staying within a monthly budget (e.g., $10,000), where they are on that spectrum, what skills they are using, and whether to use a cheaper model like Sonnet instead of Opus. Engineering managers worry about broader usage and adoption across teams, such as an observability team blowing past costs while a product team delivers fewer KPIs at lower cost. Finance folks worry about budgets, alerting, and justifying ROI on spend to their bosses, the CEO, or the board.
AI Engineer · Why No One Agrees on What Their AI Tokens Cost - Pratyusha Duvvuri, Cisco
Report · FuelThe host reports having more energy and being able to cut his caffeine content in half while drinking Ryze mushroom coffee, going from two cups of regular coffee (about 80–90 mg caffeine per cup) to two cups of mushroom coffee (about 40 mg caffeine per cup).
Boards, Bikes, And Hikes · I Tried RYZE Mushroom Coffee for 3 Months...Here's What Happened
Creator perspective · MoneyIn the observe section, it was initially seductive to just YOLO it — have Claude or some other model look at the entire corpus of things and see what it does. But it's pretty clear that a lot of information is lost; it loses attention to detail. What ultimately worked was to engineer a simple agentic workflow with a series of specialists that break down the collateral of these assets, look for specific things, and get reviewed by other agents. The amount of insightful updates to a particular project more than doubled after doing this.
AI Engineer · Your Company Has No Model of Itself. Agents Can Build One - Joao Wenzel, Citadel Securities
Creator perspective · MoneyYou can do prediction because you can look back and establish a snapshot as of some past date and essentially run a backtest. Take a snapshot as of February, look at the future labels that came after, and compare what the agent said was going to slip or flagged versus what actually happened. Statuses that were green but flagged were twice as likely to actually turn yellow or red. The system is not predictive of exactly what's going to go wrong, but it should tell you what to look at.
AI Engineer · Your Company Has No Model of Itself. Agents Can Build One - Joao Wenzel, Citadel Securities
Report · MoneyLion's mane sells for $20 per pound retail, same as oysters. Wholesale lion's mane/oysters sell for $10 per pound. Retail is preferred over wholesale because of higher profit margins. There's a trade-off with retail: you have to physically stand at the markets or pay someone else to, which takes time away from growing more mushrooms or doing other farm work. On the wholesale side, you drop off mushrooms and still make really good money — margins are around 95% even selling wholesale once equipment costs are covered.
UpFlip · Here's How to Start a $10K/Month Mushroom Business from Home
Report · MoneyThe source states you can grow oyster mushrooms in a home or backyard, and that he sells them retail direct to consumer at $20 per half pound, which is $40 a pound. The setup-to-harvest time frame is seven to ten days. Once set up, growing is described as basically taking trash and turning it into treasure.
UpFlip · Here's How to Start a $10K/Month Mushroom Business from Home
Creator perspective · MoneyAmjad Masad describes how users who coded 100% with AI would open ChatGPT in one window and Repl.it in another, copying code into Repl.it, pasting the resulting error back into ChatGPT, and copying the fix back — such a manual process that the insight was to automate the entire process. To make AI coding agents work, you have to automate the back-and-forth AND understand the errors, not just relay them; the terminal condition for the loop should be that the program is working — loop to completion rather than doing one turn. Replit demonstrated you can run LLMs to completion, and as models get more capable, the engineering needed to make them become agents decreases.
Y Combinator · How To Build The Future: Paul Graham & Amjad Masad
Creator perspectiveTheis describes his lab's typical approach: learn representations with a contrastive model (e.g., SC Concept), then apply a generative model like flow matching to push forward along the action for perturbation prediction. Generative perturbation prediction approaches are much better at generating perturbation effects but are highly dependent on a good latent space. The SelfFlow paper, accepted at Nature Methods, demonstrates that generative approaches with good latent spaces significantly outperform alternatives for perturbation effect prediction.
Latent Space · Controlling Biology - Fabian Theis, Helmholtz Munich
Creator perspectiveTheis describes the next frontier as moving beyond mapping and prediction to steering experiments: using models to guide the next experiment in a closed loop rather than ending projects at a model and concept figure. Lab Compass is an inverse generative model on top of a forward world model that tells you what experiment to run next to reach a target cell state. The ultimate vision is to run reinforcement learning in the experimental loop for biology. He also frames model papers as starting points for continual experimentation rather than ending projects at a model and concept figure.
hedged
Latent Space · Controlling Biology - Fabian Theis, Helmholtz Munich
Creator perspective · MoneyPhilipp Schmid explains that you can define how and where credentials should be used — e.g., for Slack, define that a credential is only allowed to be used with a certain domain or a certain environment variable, so you control when a credential is used. When a credential is defined as an environment variable, the model has a local dummy value; when it makes the call, the network proxy monitors it, sees the domain the user defined, swaps the dummy value with the real credential, the call succeeds, and the agent gets back the result without knowing its dummy token wasn't working. A network proxy sits around the sandbox and watches all network traffic going in and out so credential usage can be controlled.
AI Engineer · 7 Things Your AI Coworker Needs - Philipp Schmid, Google DeepMind
Report · MoneyAI allowed Freshworks to set up an agentic pipeline with human checkpoints to get more clarity about whether an idea deserves to be pursued. The pipeline automatically generates a design prompt carrying requirements and gathered context, which is used to spin up an automatically generated prototype using the design system. This prototype acts as an early checkpoint for product managers and designers to test their basic assumptions. Finally, the pipeline generates a PRD that allows building out the idea further with full context, requirements, the design prompt, and recorded decisions. At the final stage, designers use the final artifact to build the feature or idea out fully in the playground.
Figma · Config India 2026: What changed when designers started shipping code
Creator perspective · MoneyIn the Earshot demo, one layer of the system transcribes audio and chooses an action based on the transcribed word, using a very small model operating on a very small slice of audio (40-100 milliseconds) with crude frequency bins as features. A supervised model knows nothing in the beginning and can't be trusted at first, so a teacher model layer teaches the small real-time model in a supervised setup, and a second layer independently figures out when the system trusts the small model enough to follow it. Early on there is a big delay between the word and the response; the model learns faster as it hears examples more times, and by the end of training the demo got down to essentially 45 millisecond response time — reacting to voice from a single word faster than we can transcribe what we's
AI Engineer · Your Agent Doesn't Need an LLM for Every Decision - Raymond Weitekamp, OpenProse