Ads get ignored on social media but not in TLDR! Reach developers, PMs, marketers, founders and other tech leaders where they actually pay attention. Learn more about sponsorship opportunities.
OpenAI recently gave a sneak peek into its next model, named Astra. An internal version of the model was able to create breakthroughs on 10 long-standing problems in mathematics and theoretical computer science. The total number of tokens used to discover all ten solutions would have cost around $2,000 at GPT-5.6 Sol API rates. Each solution was formalized in Lean, allowing them to be verified using the theorem-proving system.
Apple's new Apple Upgrade leasing program hints at the company's next era of growth. Due to memory shortages, hardware prices have increased, and prices are only going to get steeper. Apple appears to be changing the psychology of the purchase. Instead of asking customers to justify spending more than $1,000 on a device, it's asking whether they can afford roughly a dollar a day, which is a much easier decision.
Any biotech company with an experimental drug in Montana now has a clear path to selling it to consumers. Companies can now pay $12,500 to apply to a newly established board for approval, as long as the drugs have been through preliminary testing. The drugs will be sold via experimental treatment clinics, the first of which is likely to be open by the end of the year. Patients must provide fully informed consent, and each application must be reviewed by a board that includes a Montana-certified doctor, expert scientists, and an ethicist.
Balaji Srinivasan, a tech founder, said that he believed the US was descending into anarchy, so he decided to build a new city from scratch. He started by creating a tech community in Malaysia, setting up a campus with co-working spaces, a high-end gym, and meals designed by longevity guru Bryan Johnson. Two years in, his vision is now colliding with reality. The project has been shut down over licensing issues. Within hours of the government announcement, Srinivasan signed an agreement to open a campus in Kazakhstan.
GitHub Stacked PRs break large changes into a chain of small, reviewable pull requests that build on each other. gh stack is a GitHub CLI extension for managing stacked branches and pull requests. It automates creating branches, keeping them rebased, setting correct PR base branches, and navigating between layers. gh stack offers AI agent integration.
Before AI, development costs made it rational for complex software to ship with large configuration files, extension systems, and plug-in systems. Now, the cost of creating software and making changes has dropped dramatically. Software that can be personalized doesn't need a plugin system or config file. Users can just make the changes they need themselves, with the help of agents if required.
Over 1M subscribers read TLDR AI to stay on top of the latest in AI models, research, engineering, and more. If you work in AI and want to help curate it, send your LinkedIn or resume to [email protected]!
Leopold Aschenbrenner had planned a multiday celebration for his wedding, but his fund started unraveling just when guests started to arrive. The 24-year-old investor had built a $45 billion investment powerhouse that primarily bought stocks in the AI trade. The firm borrowed too much money to make these AI bets, leaving it at risk as they faltered. Citadel then swooped in and purchased the vast majority of the firm's public stock portfolio at a discount of more than 10% to the market value of those shares. The firm is left with a portfolio of stocks and stakes in startups valued at more than $10 billion, but it is still alive.
When Trump came into office, he took the guardrails that were hindering AI development, and Larry Ellison cashed in by signing several large infrastructure deals. This big bet on AI was built on an astronomical amount of debt in every imaginable form. The hypothesis was that whoever controlled the most computing power would control the AI economy. Investors have started questioning whether this level of spending is sustainable in recent weeks and are concerned about whether the money being pumped into data centers will ever return the promised profits.
Agent traffic grew 15x last year alone. ZeroClick turns your product into an agent-purchasable service with x402 & MPP support, a machine-readable storefront, agent transaction analytics and more.
SpaceX has been hunting for a spectrum that works well in cities and dense areas for its wireless network. It is considering buying competitors to acquire the spectrum or competing at a government auction set for next year. The company has expressed interest in building its own terrestrial networks and has demonstrated a prototype mobile handset to some investors. Entering into the mobile market is capital intensive, so SpaceX could ultimately choose not to follow through on these plans.
OpenAI's annualized recurring revenue in July exceeded the company's entire second quarter. The growth was driven by the release of the company's GPT-5.6 series of models and ChatGPT Work, as well as growing adoption of Codex. OpenAI is under pressure to justify its $852 billion valuation as it prepares for its IPO. It has been racing to bring in new users to generate the revenue needed to help support its infrastructure spending plans.
DoorDash has received permission from the FAA to deliver packages by drone at altitudes below 400 feet. The drones will be built in the company's in-house robotics unit, DoorDash Labs. DoorDash Air aims to cater to any merchant, anywhere. The company will release more information later this year on its drones and how its drone program will fit into DoorDash's broader ecosystem.
China has accelerated its deployment of civilian and military hardware into orbit. The country's growing constellations of satellites could help it track events on Earth in real time and in greater detail than ever before. This has raised concerns among US officials, who see China's growing array of space-based capabilities as a quest to displace the US as the pre-eminent space power. Western analysts say that China is leading the world in research in sensors and satellite positioning technologies.
Ever feel like you're shipping AI in slow-motion? This Datadog kit includes an AI monitoring eBook, product briefs, and an on-demand webinar on software delivery best practices. Download it to get up to speed on delivering observable AI. Grab your Developer Toolkit for the AI Era
Prompts are dependencies. Changing them can shift downstream behavior with no warning, no failing build, and no obvious signal in the request pipeline. Treat prompts exactly like any other artifact that can change production behavior. Test them, give them a gate, and default that gate to blocking until proven otherwise.
AI labs are moving faster than ever. However, capability is only half of the picture. Optimizing the cost per successful task makes models more affordable and less costly to serve. This article discusses techniques adopted in frontier labs to make API applications more efficient. It covers what happens when requests are sent to an AI agent, how the harness layer cuts repeated work, and more.
A new class of entrepreneurs is launching and running new million-dollar companies on their own. AI's ability to handle various administrative tasks makes it useful for launching solo businesses in many fields. The effect of one-person businesses on the labor market remains to be seen. While many people are worried AI will limit employment opportunities, these solo operators show how the technology can also open new doors.
500,000 people have already signed up for TLDR Hardware, our new twice-weekly newsletter covering chips, robotics, energy, and devices. If you work in hardware and want to help curate it, send your LinkedIn or resume to [email protected]!
Labs do not want to spend greater and greater shares of compute on inference. The point of inference revenue is to convince investors to spend more money to buy more compute to train bigger models. Spending too much on compute on inference signals that progress has stalled because it's not worth investing more in training. Labs believe that their models will continue to get better, so this only means that compute will get more expensive.
WhiteFiber validated 111.2 Tbps across 83km of dark fiber, within 8% of the physical speed of light. Two geographically separated sites. One logical AI supercluster. See the data
Chinese founders face more pressure and liability, which results in more competent and ruthless companies, but also creates incentives for more froth and fraud.
Apple is preparing a new push into the smart home category with a wave of home products. It will start with a hub device built around the new Siri AI assistant. Apple is also preparing to release a new TV set-top box and a refreshed HomePod mini. All three devices are nearly ready to launch. Apple is also working on a higher-end, robotic version of the home hub and an advanced in-home security camera.
Claude Mythos Preview found flaws in a watered-down version of an algorithm for Advanced Encryption Standard (AES), a ubiquitous protocol that protects web traffic, wireless networks, data storage, and many other technologies. Anthropic claims the assault was 200 to 1,000 times faster than what previous human research had managed to do. The attack was devised almost entirely autonomously - the model needed aid as it originally believed the problem was impossible. Claude Mythos Preview worked on the problem for about a week before engineering its attack, and two human researchers worked for nearly a month to verify that the method appeared correct.
Tau's humanoid service is now available in San Francisco. Access is initially invite-only. The service costs $30 per hour. Each humanoid will be jointly controlled by a human operator and AI. Video of the robot in action is available in the article.
Organoids are clumps of human cells that show some type of organ-specific function. In theory, they seem like a good tool for drug discovery, but the reality is that the technology is still not reliable enough. There are many people working to fix the issues with organoids, but they are still methods in progress. Organoids have not dramatically altered medicine yet, but they have already been used for some unarguably useful things.
Every token in the orchestrator's context is competing for its attention. The real value of a subagent is what it keeps out of that context. Subagents should be treated as a tool for protecting the orchestrator's working memory, offloading reasoning it doesn't need to hold onto. Doing this well means giving the orchestrator explicit ground rules for when and how to delegate.
MCP 2026-07-28 is the largest update to the protocol since launch. It is now easier to deploy and scale remote servers. MCP is now stateless, so it can be deployed on serverless and edge infrastructure, or scaled horizontally behind any load balancer. There is now a formal path to extend the protocol.
Anthropic has drawn criticism after releasing tools on the market that compete with those offered by partner companies. For example, Claude Design directly competes with offerings from partner design company Figma. Anthropic's data retention policies have also been criticized, but the company has pledged it won't use conversation data from Fable and Mythos - but not other models - to train new models. Researchers are calling for more transparency in AI systems.
500,000 people have already signed up for TLDR Hardware, our new twice-weekly newsletter covering chips, robotics, energy, and devices. If you work in hardware and want to help curate it, send your LinkedIn or resume to [email protected]!
As AI continues to improve, the question isn't whether superintelligence will exist, but who will have access to it. The technology could be centralized and restricted to a few institutions, or it could be used to empower everyone. The notion that AI is such a threat that the only safe path is an extreme concentration of power seems dangerous. Historically, hoping that an absolute power will benevolently provide for humanity if sufficiently enlightened hasn't led to safe or positive outcomes.
You're on track to close your first enterprise customer, but they're asking about SAML, SCIM, and OIDC. This PropelAuth blog answers the 7 most common enterprise SSO questions. Read the blog
DoorDash uses LLMs offline to enrich a knowledge graph, Instacart uses them at the query understanding layer, and Uber Eats fine-tuned a model into the embedding backbone of two-tower retrieval.
Amazon plans to launch 5,105 low-Earth orbiting satellites to build on top of Globalstar's existing network of only a few dozen satellites, which it acquired in an $11.6 billion deal a few months ago. The network will power both voice and mobile data. Amazon has yet to release pricing for its service. In the US, the satellites will use Globalstar's radio spectrum in the 1.6/2.4GHz bands, and orbit between 510 km and 580 km in altitude.
OpenAI is close to leasing a $500 billion data center in southern Ohio. The company is in talks with Nvidia for a $250 billion financial backstop for the project. The US government is still talking to other potential tenants. The deal will not be final until Commerce Secretary Howard Lutnick agrees to it.
A Shanghai-based company has started mass producing immersion deep ultraviolet lithography tools. The unnamed firm aims to make about five DUV machines this year and is targeting about 20 next year. Production is still at an early stage. DUV machines are less advanced than ASML's sophisticated extreme ultraviolet lithography tools. ASML has been banned from exporting its tools to China.
Frontier LLMs were trained on text data from the internet. There's no equivalent source of physical data for robotics. All attempts to solve and get to physical AGI involve working through this data bottleneck. Progress is being made on many fronts, but nobody knows how much data is enough.
Seal is an open standard for proving any file is real. It creates sealed artifacts that can be verified by anyone. A seal proves integrity, time, and the issuing certificate. Seals are anchored to a public ledger that can be checked by anyone, forever.
Octane is a React replacement UI library with improved developer ergonomics. It is incredibly efficient and is designed to stay close to the speed of hand-written DOM code. Developers write with the React API they already know, and Octane's compiler removes the virtual DOM, Suspense waterfalls, rules-of-hooks bookkeeping, and hand-maintained dependency arrays before the app ships.
500,000 people have already signed up for TLDR Hardware, our new twice-weekly newsletter covering chips, robotics, energy, and devices. If you work in hardware and want to help curate it, send your LinkedIn or resume to [email protected]!
Nvidia has made a substantial investment into Safe Superintelligence, a company started by former OpenAI Chief Scientist Ilya Sutskever. The investment is part of a long-term partnership and will involve Safe Superintelligence receiving access to large amounts of Nvidia's flagship GPUs. Safe Superintelligence had previously been relying primarily on chips made by Google. Nvidia made the investment after getting a rare glimpse into the state of the startup's research, which has so far been kept secret.
It appears that Comcast is selling Peacock to YouTube without actually selling Peacock to YouTube. All of Peacock's content will be included in YouTube Premium subscriptions in the US starting early next year. The partnership will allow Peacock's content to be monetized through advertising. Netflix may also be exploring a Peacock add-on option.
Fine-tune your first model in 15 minutes with Crusoe Serverless Fine-Tuning. Pick from a curated library of top open models. Billing stops when they stop improving. Get started →
X is rolling out Premium and Premium+ subscribers in the US, enabling them to instantly send, receive, or request money from others on the platform without fees or limits.
Ads get ignored on social media but not in TLDR! Reach developers, PMs, marketers, founders and other tech leaders where they actually pay attention. Learn more about sponsorship opportunities.
Built by the team benchmarking every major model launch, Gray Swan simulates enterprise environments and hits them with an automated red-teaming agent trained on threats that haven't been published yet. See how secure you are →
A team based in China modified AlphaFold to help identify what parts of a gene-editing protein enable mistakes. They then modified these areas to reduce problematic off-target effects. Producing gene-editing systems tailored to prevent known off-target events is useful for fine-tuning protein-DNA interactions, which could have applications beyond gene editing.
Nvidia is in talks to provide a roughly $250 billion backstop for OpenAI as part of a massive data-center project. The guarantee will help OpenAI lease a 10-gigawatt project that Softbank's energy subsidiary is developing. Nvidia's backing allows the data center developer to raise debt at more favorable terms. Under the deal, Nvidia would guarantee a series of financing vehicles intended to make lenders feel more confident that funding behind the project is secure.
Automatic programming, when done well, means to assume the role of Linus, with the AI agents assuming the role of the different maintainers of the different subsystems.
The Isochrones API calculates the area that a person can reach from a specific origin point within a set amount of time. An isochrone accounts for actual road network travel times to create a polygon that represents true reachability. The API can be useful for delivery logistics, finding properties within a set distance from a landmark, and urban planning.
500,000 people have already signed up for TLDR Hardware, our new twice-weekly newsletter covering chips, robotics, energy, and devices. If you work in hardware and want to help curate it, send your LinkedIn or resume to [email protected]!
AI is more of an economically useful commodity than a supremely intelligent being that will make humanity subservient. Several leaders in the AI industry have recently come out in support of open-weight models. If these leaders believed that AI was going to become conscious and take over the world, they likely wouldn't be as openly pro-diffusion.
Anthropic removed over 80% of Claude Code's system from its more advanced models. This post discusses how to apply the lessons its team learned from context engineering in Claude Code. The prompt is only a small part of the context Claude gets. The rest of the context is assembled from Skills, CLAUDE.md files, memory, and other sources. Context engineering has a big impact on the results generated when using Claude Code.
Companies like Anthropic and OpenAI claim some AI models are too dangerous to be developed in the open and must be tightly controlled for safety. The rest of the industry says that open source AI models must remain open for people to further develop technologies and build new businesses. China's rapid progress in open source AI models is escalating the fight. US officials are still debating whether to approach regulating Chinese open source models as a national security issue or issue a kind of blanket ban.
SpaceX's 13th Starship test flight ended in an intact splashdown in a remote section of the Indian Ocean on Friday. The vehicle appeared to be in fine shape on landing. SpaceX will attempt to catch the ship with its launch tower on the next test flight. The next steps for Starship are getting to orbit and returning the ship to the launch site, which will give SpaceX the capability to begin launching operational Starlink V3 satellites.
Seller is an early experimental app from Meta that aims to be to sellers what tools to edit videos were to content creators. It is free and will also be available in web browsers. The app has AI features that make selling items less of a hassle. Users can sign into Seller with their Facebook accounts, and listings are automatically synced.
Get the most interesting stories in startups, tech, and programming delivered in a free daily email.
Stripe is in talks to acquire OpenRouter, a startup that helps developers choose between AI models. The talks are still ongoing and could fall apart. Industry experts estimate that the business could fetch about $10 billion in a sale. Stripe is separately pursuing a deal for PayPal.
ChatGPT Health is now available to all US-based users over 18. The feature helps users with health-related queries. Users can let the feature access data from apps and integrate medical records from hospital systems. While several studies have found that AI bots are not reliable for medical advice, this has not deterred AI companies from launching health-related features.
Unitree's GD01 is a transformable mecha robot. It can transport a pilot in its torso-mounted cockpit, walk on either two legs or four, and wield fists powerful enough to knock down walls. Standing 9 ft tall and weighing half a ton, it has heavy-duty mechanical hinges that connect titanium-alloy limbs to a protective carbon fiber casing. Pictures of the robot are available in the article.
SpaceX has started to turn away satellite operators seeking dedicated rides on its Falcon 9 rocket beyond 2028. The company has also stopped building some non-reusable components for the Falcon family. SpaceX has publicly stated that it intends to offramp its Falcon 9 rocket for Starship, but has not disclosed a timeline. The company is likely to still use the Falcon 9 for launches for the Department of Defense and NASA.
No amount of harness engineering or loopsmaxxing can solve what is fundamentally a model-training issue. Engineers are still stuck reading code. Companies can move 10x-100x faster by convincing themselves that the code doesn't matter anymore, but they could also embrace their constraints and just move 2x-3x faster, safely.
The field of formal verification and deterministic system testing is about to cross a chasm. There are not enough skilled practitioners in the field. Bug-catching tools, when used in conjunction with adversarial code reviews by an LLM and with language analyzers driven by pre-commit hooks, will be a key component in software factories. They will enable people/agents to deliver reliable software without the burden of having to learn this specialized knowledge.
500,000 people have already signed up for TLDR Hardware, our new twice-weekly newsletter covering chips, robotics, energy, and devices. If you work in hardware and want to help curate it, send your LinkedIn or resume to [email protected]!
Optimizing the cost-to-outcome frontier in real world context means using the right model for each task, and optimizing the context, skills, tools, and agent harness around it. It's possible to take saturated frontier capabilities and deliver them at scale and at lower cost through models optimized for high-usage products, while continuing to use frontier models for frontier needs. The model is only one part of the hill-climbing system. Harness, memory, context, tools, skills, and user interactions all shape the evals and performance of these agentic systems.
Huawei is leading the way in reducing China's dependence on foreign AI chips. The company claims it has figured out workarounds for making near state-of-the-art silicon without leading-edge machinery. Huawei's deputy chairman attributes the company's progress to the restrictions the US has imposed on hardware exports. China still lags far behind the US, and chip-production capacity problems are preventing Chinese companies from deploying AI as quickly as they would like.
