B a good chatbot

Thus, our conclusions: This I believe is the first demonstration of the need for hybrid search. Hence the claim that hybrid search is a @UWaterloo innovation. You're welcome! The broader lesson is that old baselines are still surprisingly important. Let's not forget them.

We raised $1M dollars to reinvent how people read. Introducing Mark II - a $159 AI bookmark. Thread below

As promised, here's a recording of my 30-min keynote and the subsequent Q&A for the inaugural late interaction retrieval (LIR) workshop, cc @bclavie @antoine_chaffin. The talk is admittedly advanced, as it's directed at an expert IR community. But hopefully still broadly useful!

Databases are arguably the most commonly used enterprise tool, and enterprises typically have many of them. Yet no popular AI agent benchmark actually tests how well agents can query, join, and make sense of data across different databases! So, we built DAB (Data Agent Benchmark): 54 queries, 12 datasets, 9 domains, and 4 database management systems, grounded in a formative study of real enterprise data agent workloads. The best frontier model only gets 38% pass@1 (across 50 trials). Lots of room for improvement!

@softwaredoug @philippemnoel Subscribed. This is the kind of content I come to x for haha

@philippemnoel Thanks! Glad you like it Yep, p99 of a single node becomes p50 of a 100 shard cluster So as you add more nodes, you will be impacted by more tail latency and need simpler retrieval primitives to reduce the per-shard variance softwaredoug.kit.com

From today's @softwaredoug newsletter: "The higher the scale, the stronger the incentive to simplify your retrieval." Couldn't agree more. I expect over time retrieval systems to shift heavily to Postgres/S3, simple, predictable dependencies every company already use.

Damn this looks so wild

My entire childhood has led me to this moment... I built AgentCraft - orchestrate your agents with your favorite RTS interface! ⚔️ Coming soon 👀

Totally agree

I answer about a dozen or so emails every week from students and early stage founders. One of the most common red flags I see are people who want to be a founder for the sake of it and are chasing ideas or guessing. It's so common I have a canned response. Here it is: (Starting the canned response here) I’m sorry to say it sounds like you’re searching for an idea. Or, you have a solution in need of a problem. Or, you just like the idea of being a founder (for whatever reason). This isn’t what you want to hear, but go get a job and work for awhile. If you have a solution that needs market validation, then work in the industry that you think that market exists. Immerse yourself in some industry, it really doesn’t matter what one, because they’re all so filled with problems that need to be solved that you can choose anything. It only takes one or two years. Then your problem isn’t going to be wondering “is this a good idea?” “What is a good idea?” Etc. The problem is going to be: which of these 10 obviously good ideas won’t be solved unless I do it, and which do I want to spend the next 10 years of my life working on? That’s the real hard question. Remember, the key questions a VC is going to ask you and you should ask yourself is: “Why this? Why now? Why you?” You should have full confidence in all of them. The easy part is confidence in all of them. Then the hard part is executing fast enough and hoping the market moves with you with external factors that are mostly out of your control. :) Don’t search for an idea. Let one come to you. Go get a job. I’m sorry to tell you that, but it’s the advice I think you need to hear. Like I said, it won’t take long, one or two years or so. But that one or two years of working is going to save you more years of your life most likely wasting your time on the easy part (finding the idea). Plus, you’ll get paid for it.

RIP Dado Banatao. We lost a truly great, once-in-a-generation Filipino technologist. I was the beneficiary of Dado’s generosity twice - first, indirectly: early in my AI journey around 2018 I got to attend a deep learning workshop held by one of his portfolio companies (@wavecomputing) and taught by one of their engineers from Silicon Valley. The second time was last June through @phildev_org where they and @oliversegovia sponsored the Philippine AI Retreat (PAIR) and gave us the opportunity to visit Silicon Valley and learn from the super talented Filipinos there. The picture I posted was from our dinner at Dado’s house in Palo Alto. Without him, these opportunities wouldn’t have been accessible to me at all and they have deeply inspired me. I learned despite my lack of background I could learn anything if I put my mind to it. Meeting Dado in the flesh along with my fellows and the talented Filipinos in the Valley made a deep impression in me that (1) we need to push the boundaries of Filipinos in tech, (2) there’s a long way to for us to go, and (3) the best thing I can personally do is to contribute to what he put in motion - for me this means building a successful Filipino tech startup. Rest in power Dado 🫡 The heights you reached and initiatives you created to uplift Filipinos in tech will always be a beacon of inspiration to me!

.@RichardSSutton, father of reinforcement learning, doesn’t think LLMs are bitter-lesson-pilled. My steel man of Richard’s position: we need some new architecture to enable continual (on-the-job) learning. And if we have continual learning, we don't need a special training phase - the agent just learns on-the-fly - like all humans, and indeed, like all animals. This new paradigm will render our current approach with LLMs obsolete. I did my best to represent the view that LLMs will function as the foundation on which this experiential learning can happen. Some sparks flew. 0:00:00 – Are LLMs a dead-end? 0:13:51 – Do humans do imitation learning? 0:23:57 – The Era of Experience 0:34:25 – Current architectures generalize poorly out of distribution 0:42:17 – Surprises in the AI field 0:47:28 – Will The Bitter Lesson still apply after AGI? 0:54:35 – Succession to AI

The keychron Q1 max is soooo nice. Makes me wanna type more

@dayvough @ycombinator Yo, nice! Thank you 🙏

@ycombinator @brbongco

Last week, we hosted a Context Engineering event in SF with @mastra and @trychroma to share how leading applied AI teams are engineering their context windows. Tune in for the talks from @jeffreyhuber, @calcsam, @Jacob_Heller, and @dex_horthy. youtube.com/playlist?list=…

AI efficiency is important. Today, Google is sharing a technical paper detailing our comprehensive methodology for measuring the environmental impact of Gemini inference. We estimate that the median Gemini Apps text prompt uses 0.24 watt-hours of energy (equivalent to watching an average TV for ~nine seconds), and consumes 0.26 milliliters of water (about five drops) — figures that are substantially lower than many public estimates. At the same time, our AI systems are becoming more efficient through research innovations and software and hardware efficiency improvements. From May 2024 to May 2025, the energy footprint of the median Gemini Apps text prompt dropped by 33x, and the total carbon footprint dropped by 44x, through a combination of model efficiency improvements, machine utilization improvements and additional clean energy procurement, all while delivering higher quality responses. See the blog or technical paper for more about our methodology and ongoing efforts. Blog: cloud.google.com/blog/products/… Link to detailed paper: services.google.com/fh/files/misc/…

🇵🇭 One of my research interests is improving the state of Filipino NLP Happy to share that we're taking a major step towards this by introducing FilBench, an LLM benchmark for Filipino! Also accepted at EMNLP Main! 🎉 Learn more: huggingface.co/blog/filbench

huh. I wonder why gpt-5 doesn't support temperature 0...

i've seen a few start-up founders panic at the sight of a well-funded competitor launching into their space. but the counter-intuitive thing is - more competition is good for you! here's why 1) competition validates the market - many other (presumably) smart hard working people recognize there's an opportunity, so you're likely onto something. in fact, the presence of 0 competitors can often be a yellow flag - you might be too early, the market might structurally suck but you just don't know it yet, etc 2) nothing galvanizes a team like a shared public enemy. it's no longer a solitary marathon, it's now a team sprint. framed well ("we deserve to win" etc), i've seen competition sharpen focus, cut the fat off roadmaps, and dramatically accelerate shipping velocity 3) your competition likely has no idea what they're doing. maybe they had a viral video launch, but half their "product demo" could have been a smoke-screen or maybe they just lost their star engineer etc - you just don't know! even the start-ups that present the best to the outside, can often be a complete circus on the inside ultimately the best thing you can do as an early-stage founder in a competitive market is to just ship ship ship. stay focused, listen to your customers, and ship great product. don't get too bogged down about the competition - they're likely going to screw up =)

It is uncomfortable but I am grateful everyday 🙏

Yes

+1 for "context engineering" over "prompt engineering". People associate prompts with short task descriptions you'd give an LLM in your day-to-day use. When in every industrial-strength LLM app, context engineering is the delicate art and science of filling the context window with just the right information for the next step. Science because doing this right involves task descriptions and explanations, few shot examples, RAG, related (possibly multimodal) data, tools, state and history, compacting... Too little or of the wrong form and the LLM doesn't have the right context for optimal performance. Too much or too irrelevant and the LLM costs might go up and performance might come down. Doing this well is highly non-trivial. And art because of the guiding intuition around LLM psychology of people spirits. On top of context engineering itself, an LLM app has to: - break up problems just right into control flows - pack the context windows just right - dispatch calls to LLMs of the right kind and capability - handle generation-verification UIUX flows - a lot more - guardrails, security, evals, parallelism, prefetching, ... So context engineering is just one small piece of an emerging thick layer of non-trivial software that coordinates individual LLM calls (and a lot more) into full LLM apps. The term "ChatGPT wrapper" is tired and really, really wrong.