TLDR 2025-01-27
DeepSeek tops app store π, Microsoft's TikTok bid π±, Apple's leaked AI memo π
Drive 526% ROI with Vanta (Sponsor)
Vanta's a game-changer for businesses and their security teamsβand we can prove it. This
new IDC study shows how Vanta drives significant ROI, saves time, and boosts team productivity.
Demonstrating trust to customers and prospects is critical to closing dealsβbut it can also be costly, time-intensive, and complex. Vanta turns GRC programs into growth drivers, all while making it easy to manage your security risks.
Get the white paper.
DeepSeek gets Silicon Valley talking (2 minute read)
Many in the tech industry have been praising DeepSeek over the recent release of its reasoning model, R1. R1 seemingly matches or beats OpenAI's o1 model on certain benchmarks. It only cost $5.6 million to train, much less than the hundreds of millions of dollars it has taken leading US companies to train their models. The company achieved all this while facing US sanctions that prohibit the sale of advanced chips to Chinese companies. DeepSeek's AI assistant is now the top free app in the Apple App Store.
Oracle and Microsoft are reportedly in talks to take over TikTok (2 minute read)
Oracle, Microsoft, and a group of investors are in talks to take over TikTok's global operations. The deal will allow ByteDance to keep a minority stake in TikTok while the app's algorithm, data collection, and software updates will be overseen by Oracle. Oracle's server network already provides the bulk of TikTok's backbone. President Trump earlier issued an executive order giving TikTok and others a 75-day reprieve from any enforcement action related to the US' requirement that ByteDance divest from TikTok.
π
Science & Futuristic Technology
What did DeepSeek figure out about reasoning with DeepSeek-R1? (4 minute read)
Chinese AI lab DeepSeek recently released a reasoning model trained on a GPU cluster a fraction the size of any of the big western AI labs that is better than the current best reasoning models. Reasoning models are able to 'think' about questions, which results in better answers. DeepSeek used a reinforcement learning approach, as opposed to a fine-tuning approach, to train its reasoning model, which resulted in cost and quality benefits. This approach can only really train the model on coding and mathematics, but this might not matter as superintelligence in coding/mathematics might transfer to other domains - this has yet to be demonstrated in practice.
What happens when we can't just build bigger AI datacenters anymore? (6 minute read)
AI's continued growth may ultimately hinge on a new kind of supercomputer that spans entire countries. This will involve stitching together existing datacenters to distribute the workload. The infrastructure required to stitch datacenters together already exists, but it is tuned more for disaster recovery and high-availability than large-scale AI training. Research is already underway to improve the technology, which will help address power challenges by allowing datacenters in different regions to work as one.
π»
Programming, Design & Data Science
Finally, an ad spot techies won't skip (Sponsor)
Tech inboxes are crowded, but TLDR stands outβand so do the ads inside. With over 5 million software devs, founders and other tech decision-makers tuning in daily, this is where ideas, tools, and products get noticed.
Learn more about advertising with TLDR.Introducing ScopeDB: Manage Data in Petabytes for An Observability Platform (12 minute read)
ScopeDB is a database built on top of S3 that unleashes the power of cloud elasticity and workload isolation. It achieves ten times cost-efficiency and performance improvement compared to databases built with shared-nothing architecture. ScopeDB servers use a single stateless binary, so there is no need to manage local disks, data sharding, or complex leader action. A demo playground to test out the features of ScopeDB is available.
Lab (GitHub Repo)
Lab allows developers to experiment with code instantly without the overhead of creating files and folders. Users can just type 'lab' with any extension (eg. lab py) and start coding. It automatically organizes experiments with smart names in a folder and cleans them up after seven days to keep things tidy. A demo video is available.
Here are Apple's two main priorities for AI this year, per leaked memo (2 minute read)
Apple has two main internal priorities for its AI efforts this year: revamp Siri and improve its AI models. Siri is undergoing a big transformation into 'LLM Siri', which will be shipping in the spring of 2026, possibly in iOS 19.4. Apple's AI models have been criticized for regularly including errors, so there is a clear need to improve them. The company has temporarily disabled summaries in iOS 18.3 for select app categories presumably until the backing AI model can be brought up to par.
The secret sauce of Chinese social media apps (8 minute read)
In the days leading up to the deadline for TikTok to be banned in the US, users flocked to other Chinese apps such as Xiaohongshu and Lemon8. These apps have many of the same features that made TikTok so appealing: an infinite scroll of hyper-targeted recommendations and interactions with friendly strangers. They blend entertainment and networking with sales to monetize their addictive algorithms. This strategy has been very profitable, making China the global leader in social commerce.
Get the most interesting stories in startups, tech, and programming delivered in a free daily email.
Join 1,600,000 readers for
one daily email