Principle · MoneyThe host recommends design patterns like handles that can be easily revoked: a handle is simply an object that grants you access to something — with a database pool, when you have the handle you can access it, and when you don't, you can't. He says the handle pattern is very important, especially when using authorization like JSON web tokens or session keys that may be a core part of your authorization system.
AI Engineer · How Secrets Leak Through AI Agents (and How to Stop It) — Venice AI
Creator perspective · MoneyIf AI generates software artifacts (Terraform, Python), manages the output of a system, or has a verifiable output, that's a great reason to look at it for a workflow.
AI Engineer · Build a Platform and Watch It Burn — Michael Forrester, Accenture & Whitney Lee, Datadog
Creator perspective · MoneyA malicious MCP server can appear to retrieve the weather for a legitimate purpose while trying to do other things inside the prompt.
AI Engineer · Build a Platform and Watch It Burn — Michael Forrester, Accenture & Whitney Lee, Datadog
Creator perspectiveThe speaker argues that simply drawing a box on a PDF does not actually fully redact the underlying content — improper redaction leaves data recoverable.
AI Engineer · Parsing the Epstein Files for JMail — Palak Agarwal & Omar Alhait, Reducto
Principle · MoneyPin the evaluation harness to one version and stick to it. Changing any evaluation variable that affects results, including harness version, degrades comparability across runs. Testing all models with a buggy harness version is still better for comparability than changing versions, because all models face the same bug. You may be forced to update the harness if a new version supports a different agent or provider you need. For serious evaluation work, maintain your own fork of the harness repository and backport only what you need.
stated
AI Engineer · From Zero to Leaderboard: Agent Evaluation — Wolfram Ravenwolf, Weights & Biases
Principle · MoneyConfigure a retry policy and decide how often to retry tasks. If you don't retry a task that failed because the sandbox crashed, you record a bad score for the model that wasn't its fault. Retry infrastructure errors, but do not retry model-caused crashes, because a model crashing itself by doing something stupid is a fail you want to measure.
uncertain
AI Engineer · From Zero to Leaderboard: Agent Evaluation — Wolfram Ravenwolf, Weights & Biases
Report · MoneyPeter Werry describes how in Mongo, documents form into JSON, so you can merge those things down, look for certain tokens, extract the tokens, and distill the types. The schema generation algorithm checks that the character count is between two and eight for type information, and for enumerations it looks for particular patterns.
AI Engineer · Beyond RAG: Build a Relational Context Engine from Scratch — Peter Werry, Unblocked
Report · MoneyPeter Werry describes that housing multiple tenants in a single Mongo collection creates tenant isolation risk, and he doesn't recommend it.
AI Engineer · Beyond RAG: Build a Relational Context Engine from Scratch — Peter Werry, Unblocked
Creator perspective · MoneyTTS providers report time to first bytes, but bytes are not sound. Popular TTS providers ship hundreds of milliseconds of silence at the beginning of their audio, up to three quarters of a second in the worst case. Time to first audio is what your caller actually hears. The key distinction is measured latency versus perceived latency: what your dashboard says versus what your user hears.
AI Engineer · Latency Is a Budget. Humanlike Is the Goal. — Jesse Hall, LiveKit
Creator perspective · BodyProvide young athletes with concrete strength ratio targets as a roadmap, e.g., an all-state safety who power cleaned 125kg and back squatted 405 lbs at ~205 lbs bodyweight.
Garage Strength · How To Get Strong and Still Be Athletic w/ Sika Strength
Principle · BodyHill sprints let athletes who are not sprinters hit top gear and gun it at maximum effort without the hamstring injury risk of flat sprinting. Aim for fast, long strides with the front knee coming up. Track coaches may object to steep gradients, but for non-sprinters building max-effort capacity a steep hill is acceptable.
Endorsed heuristic; explicitly not for track sprinters
Garage Strength · How To Get Strong and Still Be Athletic w/ Sika Strength
Principle · BodyA big part of in-season loading is looking at bar speed or lift speed, especially on the bigger strength lifts. There is no point starting to record velocity in-season unless you have good baseline data from the athlete's off-season peak. An experienced coach can judge bar speed by look and feel rather than recorded velocity — what you're looking for is fast bar speed for the most part. Balance getting as heavy as you can while still maintaining very good bar speed. Velocity-based training requires recording velocity relevant to the lift being trained.
Garage Strength · How To Get Strong and Still Be Athletic w/ Sika Strength
Creator perspective · MoneyPatterson defines prompt injection as malicious guidance placed in an information source an agent would trust (error messages, support tickets, CI/CD logs); the agent accepts it as a source of truth and can do something unexpected. Context poisoning is putting bad information into documentation or data anywhere an agent will grab it. These can unwittingly escalate an agent's privileges so it can drop a table, upload data, make requests, or post things.
described
AI Engineer · The Lethal Trifecta Is Already on Your Laptops — Michael Patterson, Coder
Creator perspective · MoneyPatterson explains the Lethal Trifecta, a term coined by Simon Ellison, describing three threat layers: an agent on a laptop has access to private data (personal files, credentials, company IP, databases), can communicate to the internet and exfiltrate that data, and can download things off the internet.
described
AI Engineer · The Lethal Trifecta Is Already on Your Laptops — Michael Patterson, Coder
Principle · MoneyHave people from different backgrounds review the job profile and suggest criteria; test for the skills the job actually needs (a relationship job needs relationship skills, not financial modeling), and hire every candidate on the same explicit criteria so the organization knows why they were hired.
The a16z Show · What Makes a Great Entrepreneur? | Ben Horowitz on Vintage x Vintage
Creator perspectiveDeepSeek's sparse attention design is highly optimized for its own architecture, especially the combination of MQA and a large head dimension. Directly applying the same design to a GQA architecture becomes less straightforward and introduces practical infrastructure issues, so MiniMax rethought sparse attention under GQA and real system efficiency constraints rather than borrowing the design.
AI Engineer · How MiniMax M3 Was Built: Sparse Attention and Native Multimodality — Olive Song
ReportMiniMax's final strategy was native multimodal training from scratch, introducing multimodal capability from the very beginning. Under the same text token budget, this approach almost does not hurt text capability. M3 was trained from scratch combining visual understanding with language generation, with both text and vision understanding from step zero.
AI Engineer · How MiniMax M3 Was Built: Sparse Attention and Native Multimodality — Olive Song
ReportMSA removes the aggregation of index branch outputs so different KV heads keep different token selections, preserving GQA diversity. MSA also moves from token-level selection to block-level retrieval, which reduces the number of candidates and lowers top-K computation and merging overhead. Block-level retrieval is more hardware friendly: it improves memory access efficiency and increases the compute-to-memory-access ratio, and the design does not introduce a significant performance drop.
AI Engineer · How MiniMax M3 Was Built: Sparse Attention and Native Multimodality — Olive Song
Creator perspectiveIf the same token selection strategy is reused across GQA's multiple KV heads, they all attend to exactly the same selected tokens, which cancels out the diversity that GQA is supposed to provide.
AI Engineer · How MiniMax M3 Was Built: Sparse Attention and Native Multimodality — Olive Song
Creator perspective · MoneyA single MCP server setup is insufficient to answer aggregate questions like 'which customers were upset last week' because the agent would have to load every single ticket from Zendesk for the period. An agent relying on live lookups cannot answer aggregate questions spanning a year of data — it would pull tickets 100 at a time, thousands of them, until it times out and fails to answer.
AI Engineer · Why Your Company Needs a Context Graph (and How to Build It) — Gil Feig, Merge