sam diego

classically under trained dream chaser. i build things

[sam diego @Vodkowski]

girls just botswana have fun

[sam diego @Vodkowski]

dying is a grave situation

[sam diego @Vodkowski]

The post claims Claude built a Polymarket quant bot that made +$221,897. The video is a neon “hacker dashboard” animation (P&L ticking from about $232k to $270k in 25 seconds), not a real Polymarket screen recording. The CTA is the usual funnel: comment + like + RT + follow so they can DM you something “free for 24 hours.”

Same template is spammed under many accounts with swapped dollar figures (@Dan1ro0, @codewithimanshu, @marryevan999, etc.). Account @david_marco45 is Blue-verified, high post volume (~44k tweets), low following — classic promo pattern. No wallet link in the post text either.

Real Polymarket wallets can show big gross numbers, but this clip and funnel don’t prove Claude did that, or that you’d get a working bot in the DM.

[Collin Rugg @CollinRugg]

JUST IN: Elon Musk unveils the Tesla Roadster by launching it off a ramp before the car flipped in midair and safely landed on the ground.

The car is a collaboration between Tesla and SpaceX.

One of the most insane things I've ever seen.

[sam diego @Vodkowski]

Anthropic shipped Claude Sonnet 5.5 — a faster, cheaper Sonnet for everyday work. The coding jump is the headline.

It runs 30%+ faster than Sonnet 5, and Anthropic says it typically costs up to 30% less per task at the same $2/$10 per MTok list price because it uses fewer tokens. On Terminal-Bench 4.0 it scores 70.6% vs Sonnet 5's 10.3%. It's the everyday complement to Opus 5.5 — bugs, docs, slides, spreadsheets — not the open-ended judgment model.

Caveat: Opus 5.5 still wins on complex, sustained judgment. And Sonnet 5.5 ships with Opus-class cyber safeguards because its cyber capability jumped — high-risk cyber tasks fall back to Sonnet 5.

https://t.co/XVis6PY8SR

[sam diego @Vodkowski]

ElevenLabs Agents can fire multiple tools in one turn.

Agent prompt config adds enable_parallel_tool_calls (boolean, default true). When it's on, supported models run several tools in the same agent step instead of serial round-trips. Same Sep 21 changelog also drops gpt-6-astra into the agent LLM list and Slack alongside PagerDuty for agent alerts.

Caveat: "supported models" only — leave the flag on by default, but verify your chosen LLM actually parallelises before you redesign a latency-sensitive flow around it.

https://t.co/DQcX2iskBk

[sam diego @Vodkowski]

Gemini Enterprise Agent Platform can turn a still into a clip — Veo (or Gemini Omni) treats your image as the first frame and generates video from a text prompt.

Supported Veo IDs include veo-3.1-generate-001, veo-3.1-fast-generate-001, veo-3.0-generate-001 / fast, plus preview lite. Media Studio: Task → Image-to-video, optional end-frame upload, aspect 16:9 or 9:16, up to 4 samples. API path is long-running predict (or Interactions for Omni) with output to Cloud Storage; input JPEG/PNG, 720p+ recommended.

Caveat: several model IDs are still labeled preview. Generation can take over a minute — use sync only when you need the bytes immediately, otherwise poll the async operation (retained up to 14 days on the Interactions path).

https://t.co/b2IOCcABeO

[sam diego @Vodkowski]

ElevenLabs launched ElevenLabs for Students — university students 18+ in the US, Canada, EU27, Australia, and UK.

Three months free across ElevenCreative, ElevenAgents, and the ElevenAPI, plus a free year of ElevenReader Ultra. Same stack students already use at hackathons; now a standing plan instead of event-only credits. Impact Program x Professors already covers 350+ professors / 1,500+ students in ~50 countries.

Caveat: age 18+, university verification, and the country list above. More regions "to follow" — not global yet.

https://t.co/GtPAsBTW9i

[sam diego @Vodkowski]

OpenAI published how Higgsfield uses GPT-6 Astra: one engineer shipping new exploration features in a day, plus ad-variation workflows for small businesses ("take my top ad and generate 100 new versions").

CEO Alex Mashrabov credits Astra's long-horizon planning and creative/engineering pairing for the speed.

Caveat: this is OpenAI's Sep 21 customer story, not a product changelog. Treat the "one day / one engineer" line as their quote — verify against your own team size and review gates before you copy the workflow.

https://t.co/fE7gwmUaTQ

[sam diego @Vodkowski]

Microsoft Foundry put voice agents into public preview inside Foundry Agent Service.

Same foundation as text agents — same API/SDK, tools, knowledge, and governance — with voice models (GPT Realtime, Azure Realtime, MAI and more), 80+ languages, and deploy paths to web, Teams, Teams Phone, and Twilio. Long-running hosted agents and tool search shipped in the same drop.

Caveat: voice agents and long-running resilience are still preview. The Fashable sample-budget cut is their customer claim, not yours until you measure it.

https://t.co/yn2062QAbz

[sam diego @Vodkowski]

most people waste tokens on "you are claude, a helpful ai created by anthropic…"

leave that blank.

what actually moves the model:

job + domain ("write PII-aware discovery notes")
format + stop rules
"say you don't know" when unsure
identity bios are attention tax. short job prompts beat brand lore.

(EMNLP 2024: 162 personas vs empty — persona didn't win on factual Qs)

[sam diego @Vodkowski]

Higgsfield published a walkthrough for running ad production end to end inside Claude — scripts alone, finished video plus TikTok post via Higgsfield MCP, paused Meta campaigns and metrics via Meta MCP.

Verified Sep 17: the finished video still uploads to Ads Manager by hand. Credits bill at standard rates through the agent (unlimited access stays on https://t.co/Gk1f6uAJ29). Free Claude only gets one custom connector, so both MCPs need a paid plan.

Caveat: this is Higgsfield's own guide. Keep explicit review gates on publish, activation, and budget — autonomy levels are their framing, not a guarantee.

https://t.co/4K65kg4MPg

[sam diego @Vodkowski]

GitHub put Copilot cloud and local sandboxes into public preview.

Local sandboxing runs agent shell work on your machine with restricted filesystem, network, and credentials — included with a Copilot seat. Cloud sandboxing runs the whole session in an ephemeral Linux box hosted by GitHub (Azure Container Apps under the hood), billed by compute, memory, and snapshot storage.

Caveat: local isolation is lighter-weight OS containment (MXC), not a full VM. Cloud needs org enablement and is still experimental in the CLI. Confirm your host supports Seatbelt / bubblewrap / BaseContainer before you depend on it.

https://t.co/E5k7OwQkyt

[sam diego @Vodkowski]

@AlexshevPm yeah. one job per ref or the model averages them into mush. identity / wardrobe / env / camera is a clean default stack — treat it like a shot list, not a moodboard dump.

[sam diego @Vodkowski]

Dreamina Seedance 2.5 ModelArk API.

BytePlus ModelArk says the official Dreamina Seedance 2.5 API is live — 4–30s video, 480/720/1080p, up to 50 multimodal references, precision edit, 10+ languages.

Token packages are on https://t.co/wN8qtrdNQD; docs sit under ModelArk. Specs line up with what Runway already exposed as a third-party model: text plus image/video/audio in, MP4/MOV out at 24fps.

Caveat: BytePlus still isn't a US-friendly path for everyone, and "API live" is not the same credit math as Runway. Check region and pricing before you rebuild a pipeline around it.

https://t.co/HVc8nrzg5f

[sam diego @Vodkowski]

Runway Gen-4.5 is live for Text to Video and Image to Video — 720p at 24 or 25fps, clips from 2 to 10 seconds, 12 credits per second on Standard and up.

Aspect ratios from 16:9 through 21:9. On Max / Unlimited (Legacy) / Enterprise, ProRes or PNG sequence adds 5 credits per second at generate time. Extra control inputs are still listed as coming soon.

Caveat: "state of the art" is Runway's line. The hard facts are the credit math, the duration cap, and what is not in the model yet. Re-run your own prompt set before you change a production default.

https://t.co/36VGeVoPmo

[sam diego @Vodkowski]

Anthropic shipped Claude Opus 5.5 (Sep 22) — first model in the Claude 5.5 family. $4 / $20 per million tokens, cache reads $0.20. They say ~40% cheaper than Opus 5 on typical workloads, and over 30% faster output.

Live on AWS, Google Cloud, Azure, and the Claude Platform as claude-opus-5-5. Same class of cyber and biology safeguards as Fable 5.1 (Life Sciences Verification for bio work). Sonnet 5.5 and Haiku 5.5 still listed as coming in the next weeks.

Caveat: the CursorBench and Terminal-Bench leads are Anthropic's. Treat the price and speed story as the hard fact; re-run your own evals before you change defaults.

https://t.co/2X2743A3Ly

[sam diego @Vodkowski]

OpenAI shipped GPT-6 Sol and GPT-6 Luna — the cost-efficiency tier under Astra.

Sol is $2 / $10 per million tokens (was $4 / $20 on 5.6 Sol). Luna is $0.10 / $0.50 (was $0.20 / $1.20). Both are live in ChatGPT Work, Codex, and the API as gpt-6-sol / gpt-6-luna. Free and Go get Luna on desktop only; Chat rollout is gradual today.

Caveat: the AutomationBench and FrontierCode wins are OpenAI's tables. Treat the 50% price cut vs 5.6 promo pricing as the hard fact, and re-check your own harness before you switch defaults.

https://t.co/LzoAgATwiN

[sam diego @Vodkowski]

AI shipped Grok 4.7 — same $2 / $6 per million tokens as 4.6, pitched as their strongest coding and knowledge-work model.

Bigger base than 4.6, longer RL on multi-hour tasks, and native Grok Bot harness training. Live today in Cursor, Grok Build, and the API. On xAI's CursorBench 4.0 table it lands at 46.3% (4.6 was 40.4%); Fable 5.1 still leads that column at 51.8%.

Caveat: those are vendor numbers. Independent boards often diverge — treat the price/speed story as the hard fact, and re-check your own harness before you switch defaults.

https://t.co/LiOGYgY4hi