Nvidia struck a reported $6B licensing deal for Poolside plus $1B in equity, and is in talks to back Perplexity above $30B. Meanwhile OpenAI dropped GPT-5.6 Sol output pricing from $30 to $20.
A reported $6 billion licensing agreement with Poolside plus $1 billion in equity and more than 100 engineers folded into Nemotron. Nvidia is also in talks to invest in Perplexity above $30 billion, up from roughly $20 billion.
Prime Intellect gave 18 frontier models the nanoGPT optimizer speedrun, targeting 3.28 validation loss on FineWeb with a 124M active-parameter model. Across 153 autonomous runs, Claude Fable 5 finished best at 2,726 steps, closing 82% of the research gap.
Treasury Secretary Scott Bessent laid out expanded secondary sanctions on Iran on August 24, promising economic asphyxiation. The same day, reports said Trump threatened a 50% tariff on cars and trucks imported from Canada.
A weekly operator deep dive, the full Dojo Pro prompt packs, and the complete prompt database. Founding members lock the launch price forever.
Join the founding list →Prime Intellect handed 18 frontier models the nanoGPT optimizer speedrun, training a 124M active-parameter model to 3.28 validation loss on FineWeb in as few steps as possible, then left them alone in sandboxes on 8xH200 machines for up to eight days per run. Across 153 autonomous runs, Claude Fable 5 finished best...
Continue reading arrow_outwardPrime Intellect's nanoGPT speedrun benchmark put 18 frontier models through 153 autonomous optimizer runs on 8xH200 boxes, and Claude Fable 5 closed 82% of the research gap at 2,726 steps.
Treasury Secretary Scott Bessent on August 24 promised to sever every economic lifeline sustaining Tehran, warning of dire consequences for countries that decline to join the expanded secondary sanctions campaign.
The S&P 500 and Nasdaq both closed lower August 24 on a technology selloff and Iran risk pricing, yet sentiment reads Greed, a gap worth watching since hyperscaler capex plans lean on those equity valuations.
Nvidia is buying model teams and search equity at the same moment Trump threatens 50% tariffs on Canadian vehicles, which are now chip- and software-dependent: the same silicon supply chain is being consolidated upstream and taxed downstream.
A reported $6 billion licensing deal for Poolside plus $1 billion in equity and over 100 engineers moved into Nemotron, alongside talks to invest in Perplexity above a $30 billion valuation. Nvidia is no longer just selling compute to model labs; it is acquiring their models, their people, and stakes in their applications. The reported plan to raise prices on some products makes the funding source explicit.
Prime Intellect's nanoGPT speedrun gave models up to eight days alone in sandboxes to beat a concrete target: 3.28 validation loss on FineWeb with a 124M active-parameter model, in as few steps as possible. Claude Fable 5's 2,726 steps closing 82% of the research gap is the first number that maps long-horizon autonomy onto real research work rather than a static test set.
OpenAI dropped GPT-5.6 Sol output pricing from $30 to $20 per million tokens through at least November 21, while Thomson Reuters launched a frontier model built on proprietary legal, tax, and news corpora rather than general capability. General intelligence is deflating toward commodity; the defensible position is shifting to whoever owns the data the model cannot scrape.
The Deliberate Two-Track Squeeze
Read full analysis arrow_outward
Deep Dive
8.1/ 10
The four pieces you were duct-taping together to make an agent actually do work are now GA on one API, and one of them drives web apps by page structure instead of guessing pixel coordinates.
This is for engineering teams who already have a concrete, boring automation problem: an internal web tool with no API, a document pipeline, a procedure nobody has written down properly. The move from preview to GA matters because you can now put these in a production path without betting on a deprecation notice, and Browser Use targeting page elements rather than screen coordinates is the difference between a demo and a workflow that survives a layout change. Adopt it for one narrow, auditable task first; the tooling is ready, your permissions model probably is not.
Anthropic updated Claude to interact more directly with Gmail and Google Drive, which shifts it from a chat window into something that touches live company documents and inboxes. Test it on a bounded job like drafting replies from one labelled thread or summarising a single Drive folder, not your whole account. The useful discipline is treating access scope as a design decision, not a checkbox during setup.
Try it arrow_outwardIt gives customers a simple channel to submit feedback and feature requests, then organises them so product decisions come from stated demand rather than the loudest recent email. The launch listing is free, which makes it a low-cost test for a small SaaS team currently tracking requests in a spreadsheet. Point it at one product surface for a month and compare what gets requested against what your roadmap already assumed.
Try it arrow_outwardIt retouches images directly in a browser tab: strip out defects, delete unwanted objects, or clear a background for a product shot, with no install and no account ceremony. Being open-source matters if you care where the image goes, and it holds a 5.0 rating across 19 reviews on its launch page. Use it for the unglamorous volume work, product photos and screenshots, and keep your heavier editor for hero images.
Try it arrow_outwardThe integration pulls dynamic HubSpot fields into HeyGen videos and embeds the output across emails and workflows, which means personalised video at list scale rather than one recording at a time. It requires payment, so the honest first test is a single segment with a clear control group measuring reply rate, not a full campaign rebuild. If personalised video does not beat plain text for that segment, you have saved yourself a production pipeline.
Try it arrow_outwardReprice every Sol call today. GPT-5.6 Sol input dropped from $5 to $4 per million tokens, output from $30 to $20, and cached input from $0.50 to $0.40. Rates hold through at least November 21, so rerun your unit economics before the quarter closes.
Audit Claude's workplace reach before you ship it. Anthropic moved Computer Use, Skills API, and Files API to general availability and added Browser Use. With expanded Gmail and Drive access, decide what the agent can read first and scope credentials per task.
Make your agent tool calls deterministic. NVIDIA NemoClaw v0.0.114 shipped August 23 with deterministic read-only MCP tool calls for LangChain Deep Agents Code plus per-stage launch-readiness probe timing. Wire the probe timings into your CI to catch regressions.
Nvidia is no longer just selling the shovels: a $6 billion Poolside licensing deal, $1 billion of equity, 100-plus engineers, and talks to back Perplexity at over $30 billion put it inside the model layer. OpenAI answered with price, cutting Sol output by a third and holding those rates through November 21, which pushes differentiation toward data and distribution. Thomson Reuters showed the other route, launching a frontier model built on proprietary legal, tax, and news corpora instead of chasing general capability. The nanoGPT speedrun results say agents can now run eight days alone and close 82% of the research gap, so build for cheaper tokens, tighter permissions, and moats you actually own.
AI analysis, world news, markets, and tools. One briefing, delivered free.