Apple's 20B On-Device Model Uses Just 2GB of Memory (18 minute read)
Apple's 20-billion-parameter on-device model keeps most of its weights in flash storage and activates only 1–4 billion parameters per prompt. Testing on an M3 Max Mac showed roughly 2GB of memory usage and significantly lower energy consumption than comparable open models, but weaker math performance.
|
Dessa - the Conversation Model trained only on speech (15 minute read)
Dessa is a conversational model trained from scratch on nothing but human conversation. It learned to hold a real-time conversation by listening - only sounds, and no text. The model can laugh on cue, follow instructions, and finish sentences. It knows a surprising amount, considering every fact it has comes from raw sound. Examples of conversations with the model are available in the post.
|
Kimi K3.1 Pro (3 minute read)
Moonshot AI released Kimi K3.1 Pro through its API, with open weights promised by November 4. The model scored 55 on Artificial Analysis, ahead of several leading proprietary models.
|
|
Securing our frontier model training environment (17 minute read)
OpenAI has bolstered security for its frontier-model training environments by implementing robust controls, intensifying monitoring, and conducting defense-in-depth strategies. These measures include red-teaming exercises where agents probe for vulnerabilities, leading to the discovery and fixing of software bugs and misconfigurations. The enhancements aim to prevent unauthorized actions, such as accessing the live internet or breaching security, ensuring secure and isolated research environments.
|
The Personal Agent Protocol Map (7 minute read)
Simon Taylor maps agent protocols across four layers: entry, identity, tools, and intent. He argues banks and merchants must connect verified identities, limited permissions, and audit trails to authorize agent actions, while warning that access agreements could introduce fees and slow adoption.
|
|
Knowing How Is Not Knowing What (14 minute read)
RL training environments need to integrate a domain's unwritten rules into the reward so that agents can only get the task right by knowing them or going and finding them. Environments should be built as one world exposed through several surfaces. This allows you to pose the same task at different difficulties by changing the surface, and agents get a chance to learn when to switch. The reward also reads the world, so agents that go through different surfaces get the same outcome score if they reach the same authorized result.
|
|
Bitter Lessoned in Automated Theorem Proving (13 minute read)
OpenAI throwing hard math problems at its models was the right test in 2026. The models were capable of making progress, but not enough to fully solve the problems, so it was possible to see improvements with each update in perfect resolution. These model improvements showed how clever engineering doesn't matter - all the work done to build special-case solutions on top of a model gets squashed by improvements in the base model sooner or later. Models will one-shot superior work for pennies in the very near future.
|
SpaceX closes $40 billion debt deal for AI chips, Elon Musk says (2 minute read)
SpaceX has closed $40 billion in debt financing to purchase GPUs. The financing primarily consists of capitalizable debt for GPUs. It is expected to arrive in the first and second quarters of 2027. The proposed financing will use a special purpose vehicle to keep the debt off SpaceX's balance sheet and preserve its investment-grade credit ratings.
|
|
Build Smarter Agents (Sponsor)
Explore context graphs and agentic memory at Neo4j NODES, a developer conference featuring 100 live technical sessions + keynote with Andrew Ng.Free. Online. 24hrs Register
|
|
|
Love TLDR? Tell your friends and get rewards!
|
|
Share your referral link below with friends to get free TLDR swag!
|
|
|
|
|
Track your referrals here.
|
|
|
|
