BUILD ⚔️ SHIP
@build__ship
Followers
19
Following
29
Media
5
Statuses
25
Joined May 2025
openai dropped a practical guide to building agents. 32 pages technical manual on use cases, orchestration patterns, best practices. pdf 👇🏼. key insights 👇🏼
1
1
11
tldr. * agents can reason through ambiguity.* they can use tools and interact with external systems .* foundation: models + tools + instructions.* arch: prefer single agent, split only when needed .* safety: through multiple guardrail layers .* human-in-loop for high-risk actions.
0
0
2
layer multiple guardrails for resilience:. * relevance classifier (stays on topic).* safety classifier (blocks jailbreaks).* PII filter (protects private data).* moderation API (prevents harmful content).* tool safeguards (risk-based execution).
1
0
3
two agentic system architectures:. → single-agent (good for start). * one model + tools + instructions.* use prompt templates for complexity. → multi-agent (scale up). * manager pattern: central agent coordinates others.* agents hand off to each other.
1
0
3
what makes something an agent?. 1. uses LLM to manage workflow execution.2. recognizes when tasks are complete.3. can correct itself.4. can interact with external systems.5. operates within defined guardrails. *not an agent*: chatbots, single-turn LLMs, classifiers.
1
0
3
Black Forest Labs’ killer response to GPT-image.
Today we're releasing FLUX.1 Kontext - a suite of generative flow matching models that allow you to generate and edit images. Unlike traditional text-to-image models, Kontext understands both text AND images as input, enabling true in-context generation and editing.
0
0
4
RT @mitte_ai: Mixture-of-Transformers (MoTs) gain traction in new model designs. Here’s a visual breakdown of how it works 🧠👇 https://t.co….
0
3
0
all in one unified model:. * beats flux.1-dev on generation benchmarks.* does free-form image editing.* "thinks" through complex visual tasks with chain-of-thought. code + checkpoints dropping soon.
1
0
0
trained on trillions of interleaved tokens. not just image-text pairs. sequences mixing text, images, video frames. this is why it can do complex reasoning and world modeling. like predicting future frames or navigating 3d spaces.
1
0
1
the architecture is wild. mixture-of-transformers design:. * understanding expert (processes text + visual tokens).* generation expert (handles image creation). they talk to each other through shared attention. no compression bottlenecks.
1
0
0
new multimodal model: BAGEL. 7b model that understands AND generates images. open source. why aren't more people talking about this?. HF:
1
0
5
Claude Sonnet 4 beats Opus on SWE-bench at 72.7%:. * Parallel tool execution during reasoning.* Extended thinking with web search/code exec.* 65% fewer shortcuts vs 3.7 (same as Opus).* GitHub using it as base model for new Copilot agent.
1
0
4
Claude Opus 4 is the best overall model out right now:. • 72.5% on SWE-bench.• Creates memory files to maintain context.• Handles multi-step reasoning + long eval chains.• Fewer hallucinations, better adherence to instructions.
0
1
5
checking out Google’s new Gemini Diffusion model:. * not autoregressive — uses diffusion for text.* generates full outputs by refining noise.* crazy fast: 857 tokens/sec.* excels at editing/code/math tasks.* likely uses transformer backbone.
0
1
4
RT @DominiqueCAPaul: I'm training a robotic arm to play Chess. Part of it is reading the board to feed it into a chess engine. I trained….
0
1
0
RT @yusufhilmi_: created a runway calculator for everyone going into the YC interviews, good luck in both cases!.link in the thread https:/….
0
5
0