Built an agent that hill climbs YouTube thumbnails. You can't publish slop and wait for analytics to tell you it was slop. Review has to happen before you publish. But a human can't review every iteration either. That wouldn't be very agentic. So spend the human effort on teaching a judge agent up front. Calibrate the evals once, and the agent can hill-climb without you watching every step. Here's how it works: > nano-banana-pro on fal generates a thumbnail for a HubSpot-style video > A Gemini 2.5 Flash judge scores it on a 5-criterion rubric > The agent reads the judge's notes, writes the next edit, and creates another > Over 4 iterations the score climbs from 4.8 to 8.8 Then a human reviews it. Their feedback becomes 5 rule cards in the judge: > The human-encoded judge re-scores the 8.8 thumbnail to 1.2 > The agent climbs again on its own to 8.4, then 10 > The human approves An LLM judge is only as good as the expertise you give it. The judge didn't need more compute. It needed taste. Calibrate your judge up front, then let your agent hill climb. Self-verifiability is what makes coding agents so good. Marketing is more subjective though, so you need different verification methods.

This prompt works best with Nano Banana 2: Prompt: Make a photo of a cute ball-shaped version of a [animal], with big adorable over-sized eyes, it has no head, instead its face is part of the ball, its head and body are one. They are in their usual habitat. Must be almost spherical and cartoonishly cute. Its eyes are looking up in the most adorable cheeky fashion.

$GOOGL LAUNCHES “PICS,” AN AI-POWERED CANVA RIVAL FOR WORKSPACE Google is rolling out Google Pics, a new image generation and editing app built around Gemini and Nano Banana. Pics lets users: • Generate and edit images with prompts • Select and modify individual objects or text • Translate and reformat text inside images • Upscale images to 2K or 4K • Resize designs for social, web, print and digital • Collaborate with teammates Pics will work as a standalone Workspace app and inside Docs and Slides, with Drive integration coming later. The product puts Google more directly against Canva and Adobe Express for business design workflows.

Here’s this week’s shipping recap 👇 — Nano Banana 2 & Nano Banana Pro are now GA and available via the Gemini Enterprise Agent Platform, Gemini API, and in @GoogleAIStudio —Co-Scientist, our new multi-agent system for structured scientific thinking, generates and refines novel hypotheses to solve complex scientific problems — dreambeans from @GoogleLabs works overnight to curate a personalized daily collection of topics that are relevant to you based on your connected Google apps — @GoogleGemma 4 12B, our unified encoder-free model, brings powerful multimodal intelligence straight to your laptop fully offline — Gemma 4 models and their drafters are now optimized with Quantization-Aware Training (QAT) to cut memory requirements and maximize on-device performance — @GoogleMagenta RealTime 2 is our open-weights live music model that you can play like an instrument using a MIDI keyboard, text prompts, and gestures

Underappreciated how OpenAI dropped GPT-Image 2 and literally no competing model has been able to decisively beat it even 6 months later! Had some hope for Google but fumbled w/ Nano Banana 2.1, though a bigger image model is in the works at GDM coming soon(tm)

I just published my overview of the NVIDIA Nemotron series (focusing on post-training)... You can read it here: Among open models / resources, Nemotron is especially useful due to its transparency. Releases often include detailed tech reports, code, training recipes, and even data in some cases. There are many minute details involved in learning how to post-train LLMs at scale, but they are hard to learn without detailed accounts from teams that have actually trained frontier-scale models. NVIDIA Nemotron provides us with a rare opportunity to study these details directly and gain an understanding of how LLM post-training actually works. The overview covers the following, which encompasses the majority of recent Nemotron tech reports: - Llama-Nemotron: - AceReason-Nemotron: - Nemotron-Cascade: - Nemotron-Cascade 2: - Nemotron 3 Nano: - Nemotron 3 Super: - Nemotron 3 Ultra: Through these works, we can see a clear progression in post-training methodologies. Relatively simple pipelines slowly transform to become more agentic, and new components (e.g., MOPD and async / long horizon RL) are introduced over time. Beyond these high-level trends, we can see all of the small details that are useful for getting the most performance out of your model; e.g., curriculum learning, multiple training stages, synthetic data recipes, data filtering / curation / mixing, balancing reasoning efforts, harness exposure, and more. I highly recommend reading this blog (and the associated tech reports) for those who want a realistic view into the practical details of LLM post-training.

$ASML $DRAM $SKHY $MU Samsung is deepening its ASML tie-up to put High-NA EUV on next-gen DRAM first, targeting 2028, to lock in process lead and push HBM performance. “The better the DRAM, the better the HBM. This is about shrinking the 10nm-class process further for ultra-fine patterning with ASML. Technically it is a big deal.” “China will use every method to close the gap. The most important tool to widen that gap again and stay ahead of rivals is next-generation lithography.” -Lee Jong-hwan, Sangmyung University Tool cost ~$370M, but fewer multipatterning steps can cut mass-production cost if yield works. My take is.... ASML's EUV tools are irreplaceable and brutally expensive. The offset is multipatterning reduction, fewer process steps per layer, which drives production cost down over time. But mastering it takes years. This is nano-level manufacturing, not your average physics. And China being locked out of these tools is a massive disadvantage. The gap with Memory Trio only widens from here.

# Elasticsearch Features and Practical Usage 🌊 The era of manually rotating ` is over. With data streams your app just keeps writing to the same name, and all the backing-index management is handled for you. 🏷️ Title: Data stream (logical stream for append-only time-series data) 🔗 URL: 📘 Overview A data stream is an abstraction layer that lets you address a set of indices, optimized for append-only time-series data, under a single name. It suits continuously flowing data like logs, events, and metrics — you read and write to one resource name without worrying about the hidden "backing indices" behind it. ⚙️ How It Works ・A data stream is made of multiple auto-generated, hidden backing indices. The naming is `.ds--<>-`, where the generation is a six-digit, zero-padded integer starting at `000001` (e.g. `.ds-logs-myapp-prod-2026.06.10-000001`). ・Only the most recent "write index" accepts new documents. You cannot write directly to older backing indices. ・Searches automatically route to all backing indices, so queries span the entire dataset. ・Every document needs an `@timestamp` field mapped as `date` or `date_nanos`. If the template omits it, a default `date` mapping is applied automatically. ・A matching index template containing a `data_stream` definition is mandatory; it holds the mappings, settings, and lifecycle policy. One template can be shared across multiple data streams. ・When an age or size threshold is hit, a "rollover" creates a new backing index and switches the write index. This is automated via ILM or data stream lifecycle. 🛠️ Practical Usage First, define an index template that includes `data_stream`. `PUT _index_template/logs-myapp-template` `{ "index_patterns": ["logs-myapp-*"], "data_stream": {}, "template": { "mappings": { "properties": { "@timestamp": { "type": "date" } } } } }` Then your app simply keeps writing to the same name. `POST logs-myapp-prod/_doc` `{ "@timestamp": "2026-06-10T09:00:00Z", "level": "INFO", "message": "started" }` 💡 Use Cases A common setup is writing application logs to a `logs-myapp-prod` data stream and letting backing-index rollover be managed automatically. The app no longer rotates date-based indices (` itself; it just POSTs to the same endpoint every time. ⚠️ Caveats ・Backing index names are an internal implementation detail. They can change during restore or shrink, so never build logic (like dates) off the names. ・It is a poor fit for frequently overwriting the same ID (last-write-wins); for that, use a regular index or an alias. ・Use dedicated APIs like `update by query` and `delete by query` for modifications, and remember you cannot write directly to anything but the write index. ・You cannot delete an index template that is in use by a data stream. #Elasticsearch# #DataStreams#

Do more than just generate an image in Google Flow. Edit, refine, adjust angles, add formatted text, and then bring it all to life with video. A quick breakdown of the workflow: - Nano Banana 2: Fast concept generation and style testing. - Nano Banana Pro: Precision scene edits, multi-angle shots, and typography. - Omni Flash: Use your image as an ingredient or frame and make it come to life.