Grok Bot 업무 자동화, DeepSeek Flash 가격 충격, 다중 모델 에이전트 API와 효율적 모델 구조

Grok Bot is hands down one of the best products I have ever used. Maybe the best ever. I have had it less than a week. I am not a developer. I talk to one chat and it already knows my business better than me. It reads Slack, email, meeting notes, Notion, and Stripe. It knows Farzad, my clients, the courses we sell, the agency, the sales calls, and who owes each step in our workflow. I do not brief it every morning. It briefs me. Course scripts go through a Claude project we linked. Claude writes. This thing briefs it, fact-checks, and makes it sound like me before I film. YouTube is another bit. It builds the 1080p slides to go in videos, zips them with a cue sheet, and drops the pack in Slack for the editors the same turn. Caught a bad number in a SpaceX video the night I was filming, rebuilt the slides, and sent the pack. Script change, slide change. I did not have to ask twice. It sat on our sales calls and reviewed the process. Then it wrote a rule from that. Book the next call before you hang up. Send the Stripe link while they are still there. Stop emailing homework. It built a live 50% off code on Stripe for an AI course and set when it dies. Stripe is a minefield when you don’t know what you’re doing, could have taken me an hour and it did it in one click. I told it my rules for life. Question every requirement. Delete before you simplify. It holds me to that on the work, without me invoking it. This is a chief of staff I hired on a Tuesday. Don’t sleep on GrokBot.

THREAD: A roofing contractor is using Grok @Bot to run his back office. A plumber uses it to book jobs. Silicon Valley heavyweights call it a breakthrough. Each agent gets a computer and actually does the work. https:// x.com/naval/status/2 090497355649008059 … Here’s what people are using it for:

Jay

@jayair

Okay let me tell you about what's happening with DeepSeek v4 Flash. First some background, it launched on Aug 1st and within 2 weeks it went from doing 3T tokens a day to 18T tokens a day on OpenCode; an absurd 6x increase. To put it in context, that's close to doubling up all of OpenRouter's daily volume. It's also likely 30-50% of DeepSeek's total volume. Jumps like these over a 2-week span are not normal. This happened because the model is absurdly cheap; 350x cheaper than Fable, 175x cheaper than 5.6 Sol, 70x cheaper than Sonnet, and 7x cheaper than Luna. And secondly, it was a marked improvement over the previous Flash model. For the first time our users got a feel for AI that's "too cheap to meter". Then on August 16th DeepSeek raised prices by 5x (for peak hours, 2.5x off-peak hours). And it completely killed the growth of the model. It's doing less than half the tokens per day from its peak. Obviously people were unhappy with the sudden change. Our guess is that DeepSeek genuinely could not handle the absurd 6x jump. Also, it's likely that the increase in GPU prices meant that even if they acquired new capacity, they wouldn't be able to serve it at the original prices. A quick aside on why DeepSeek Flash is so cheap. It looks like they are running some custom infrastructure to cache way more tokens, for far longer. This matters because we've been scrambling trying to find providers that can fill this near 10T token per day void left by DeepSeek Flash. Unfortunately there are just a couple of people who are able to match DeepSeek's original pricing and that's likely only the case because they are using newer hardware. That brings us to the current state of things. Over the last week we've talked to as many people as possible to get DeepSeek hosted at the original price. The issue is that even if somebody is able to, it's very hard for them to have enough capacity to handle our volume. It'll take roughly 1000 B300s to handle our throughput. This is why if you've been using DeepSeek Flash on Go over the past few days, you might not have had the best experience. We've unfortunately cycled through a few different providers. This 10T token per day gap, though, is an opportunity for every other model lab. It's very clear there's an appetite for a model that's at least as competent and cheap as DeepSeek Flash. And somehow that still feels like the floor.

Aravind Srinivas

@AravSrinivas

A full-fledged developer platform for AI should provide you with access to various models (frontier and workhorse), as well as tools for deploying them in useful production workloads. That is basically the Perplexity Agent API.

"Matryoshka Language Model Suites" Instead of training every model size separately, this paper nests 500M, 1.5B, and 3B models inside one architecture and trains them together. The smaller models are standalone checkpoints, get near-free distillation from the largest model, and share weights + KV cache for speculative decoding. And you still get the same performance, with 36% less training compute, and 14-26% faster speculative decoding. So a model family can become one jointly trained system instead of several independent models. https:// alphaxiv.org/abs/2608.09703

"The Embedder's Dilemma: LLMs Are Better, but at What Cost?" LLMs now match the best embedding models overall, but at up to 1,431x the cost. Embeddings still win on classification and tie on similarity and clustering, while LLMs are better on reasoning-heavy retrieval. So a good decision maker they proposed is use embeddings by default, and spend LLM compute only where reasoning actually helps. https:// alphaxiv.org/abs/2608.12875

Finally a good paper testing whether memory-based self-improving agents actually improve. The re-evaluation adds two things prior work skipped, multiple runs to measure variance and randomly shuffled task orders. Both hurt. Agent evaluation is already noisy on multi-step tasks, and stacking a self-improvement loop amplifies that noise. Default task orderings impose an implicit curriculum that much of the reported gain was riding on. Adding detailed rubrics and environment feedback to memory construction recovers part of the drop, and a significant gap remains. Paper: https:// arxiv.org/abs/2608.18066 Track more trending AI papers in our academy: https:// academy.dair.ai

Leverage a PyTorch-native fine-tuning library with Day-0 Hugging Face checkpoint support for the recently released Alibaba open weights for Qwen3.8-2.4T-A95B (Qwen3.8-Max), its largest open-weight model. Developers can post-train the model for domain-specific use cases with configurable reasoning using PyTorch-native tools and NVIDIA NeMo AutoModel. Train directly on existing checkpoints without model conversion, with support for full SFT or memory-efficient LoRA fine-tuning. Read the full post: https:// bit.ly/4xdN25l

Ever wanted someone else to do your texting for you? ChatGPT is being offered up as an automated text scribe via a new Apple Messages integration.