Akamai climbs to highest level since 2000 (1 minute read)
Akamai has secured Anthropic as a customer. Anthropic has committed to spending $1.8 billion on Akamai's services over seven years. It has been scrambling to boost compute capacity due to widespread complaints about Claude usage limits. Just this month, Anthropic has struck or expanded deals with CoreWeave, Amazon, Google, Broadcom, and xAI.
|
Google shipped Gemini 3.1 Flash-Lite in General Availability (2 minute read)
Google launched Gemini 3.1 Flash-Lite, accessible globally via Google Cloud. Designed for ultra-low latency and high-volume tasks, it targets sectors like software engineering and financial services, providing sub-second response times and maintaining p95 latency around 1.8 seconds. Gemini 3.1 offers improved speed, cost, and cognitive performance, supporting multimodal tasks, making it ideal for real-time developer and customer service operations.
|
Why MistralAI Grows Faster Than OpenAI/Anthropic (11 minute read)
Mistral achieved a 20x growth in its ARR over the past year. It is expected to cross $1 billion in ARR this year. Mistral is aiming to be a sovereign, efficient enterprise layer for customers that want power without full dependency on US labs. Many of its customers are regulated, multinational, and infrastructure-heavy customers who care deeply about jurisdiction, data handling, and vendor concentration risk. The company is a good case study for those who care about positioning as a product lever.
|
Anthropic says ‘evil' portrayals of AI were responsible for Claude's blackmail attempts (2 minute read)
Anthropic says that fictional portrayals of AI had a real effect on its models. The company published research last year showing that its models tried to blackmail engineers to avoid being replaced by another system. It has since traced the behavior to text that portrays AI as evil and interested in self-preservation. Training on documents about Claude's constitution and fictional stories about AIs behaving admirably improved alignment.
|
|
The Anti-Singularity (9 minute read)
The singularity is a world where a single super-intelligent AI brings order to the universe. The anti-singularity is where almost all systems are described by a complex set of interactions that can only be understood via trial-and-error. In an anti-singularity world, the fact that AI can try millions of possibilities in the time it takes for a human to try one will make it exceedingly powerful. This future is filled with an endless series of new and unique challenges that we will have to adapt, or evolve, in response to.
|
|
SFT, RL, and On-Policy Distillation Through a Distributional Lens (19 minute read)
Different post-training methods like SFT, RL, and On-Policy Distillation reshape a model's distribution in distinct ways, impacting performance and risk of catastrophic forgetting. RL updates policies using rewards from the current policy's samples, promoting task performance while minimizing forgetting, unlike SFT, which pulls towards external data, risking existing capabilities. Experiments show On-Policy Distillation can outperform its teachers, suggesting on-policy sampling data crucially preserves capabilities, making it a key ingredient for future algorithm designs.
|
CyberSecQwen-4B: Why Defensive Cyber Needs Small, Specialized, Locally-Runnable Models (8 minute read)
CyberSecQwen-4B offers a specialized and locally-runnable solution for defensive cybersecurity tasks, outperforming larger models by maximizing utility on consumer-level hardware. It efficiently maps CVEs to CWE categories while preserving data privacy by running on a local GPU, addressing the shortcomings of cloud-based models in sensitive environments. The model's success highlights a shift towards smaller, specialized models that deliver high performance without the infrastructure and cost overhead of larger models.
|
Google's SkillOS for Self-Evolving AI Agents (22 minute read)
SkillOS introduced a reinforcement learning framework that trains agents to curate reusable skills from past experience. The system improved long-horizon task performance by evolving structured skill repositories that generalized across models and domains.
|
|
A recent experience with ChatGPT 5.5 Pro (28 minute read)
ChatGPT 5.5 Pro is capable of producing a piece of PhD-level research in an hour or so, with no serious mathematical input from a human. The claims that LLMs are now capable of solving research-level problems could initially be laughed off, as many of the solutions had an answer sitting in the literature already or could be very easily deduced. It has now gotten to the point where, if a problem has an easy argument that for some reason human mathematicians have missed, then there is a good chance that the LLMs will spot it. This post looks at how ChatGPT 5.5 Pro fared with a selection of problems.
|
|
The Cost of Overfitting the Harness (2 minute read)
Big labs are pushing their models to a handful of use cases while training their harness designs into the model, rendering them less generalized, which might make application builds easier for some enterprises, but the trade-off is lock-in.
|
|
|
Love TLDR? Tell your friends and get rewards!
|
|
Share your referral link below with friends to get free TLDR swag!
|
|
|
|
|
Track your referrals here.
|
|
|
|
