Thomas Housset

AI & Cybersecurity engineer.
building

[Thomas Housset @houssethomas]

https://t.co/745Mx1NdJI is live.

The Wayback Machine remembers everything a website ever published. 400,000 archived pages for a single domain, and existing tools only hand you the list of URLs.

So I built the one that reads them: every capture analyzed, e-mails, subdomains, API keys, trackers and wallets extracted, each finding dated and traceable to its exact capture.

Thousands of hours of manual review, in one run.

Free, open source.

[Thomas Housset @houssethomas]

This is how you promote a model: doing literally anything but staring at it while it works

[Thomas Housset @houssethomas]

@Alibaba_Qwen This is how you promote a model: doing literally anything but staring at it while it works

[Thomas Housset @houssethomas]

@arena @Alibaba_Qwen Qwen over Fable wasn't on my 2026 bingo card

[Thomas Housset @houssethomas]

@buildanythingso i’m in

[Thomas Housset @houssethomas]

@theo Just tried T3 and it’s amazing, keep up the good work !

[Thomas Housset @houssethomas]

@shiri_shh @boltdotnew Useful, but not new: Bolt’s release notes date automatic publish-time reviews to October 2025. Its docs say the tool flags issues and offers “Ask Bolt to fix”; they do not say publishing is blocked. Better than “later,” but not an independent security review.

[Thomas Housset @houssethomas]

AI now has express shipping.

OpenAI cut Luna’s price by 80%. Sol Fast costs 2x for up to 2.5x the speed.

When either model can do the job, route by what hurts more: the bill or the clock.

[Thomas Housset @houssethomas]

OpenAI did not cut your ChatGPT or Codex subscription price. It made the same plan last longer.

Quota budgets stay unchanged, but Luna and Terra now consume fewer credits.

Same monthly bill. More miles before empty.

[Thomas Housset @houssethomas]

@kimmonismus One caveat: Dreamina’s own Seedance 2.5 page still says “Coming Soon,” even while listing 30-second generation, a 180-second beta mode and 50 references. The specs are public. Full availability is less clear.

[Thomas Housset @houssethomas]

@aisearchio The price gap is real, but “100x” depends on the token mix. DeepSeek lists $0.14/M input and $0.28/M output; Opus 4.8 is $5/$25, so the gap is 36x to 89x. Artificial Analysis also scores V4 Flash 0731 at 50 vs 56 for Opus. Huge deal. Not blanket parity.

[Thomas Housset @houssethomas]

@nutlope The hidden number is $0.003/M cached input, 98% below the normal input rate. For agents that resend the same system prompt, tool schemas and repo context, cache hit rate may matter more than the headline $0.14.

[Thomas Housset @houssethomas]

@jun_song DeepSeek’s V4 Flash is 284B total, with 13B activated per token. That 13B describes MoE routing, not model size. The efficiency story may be huge. “Haiku-sized” doesn’t follow from these numbers.

[Thomas Housset @houssethomas]

@Da7_Tech The price war is real, but Luna hasn’t matched DeepSeek on output. OpenAI cut Luna to $0.20/M input and $1.20/M output; V4 Flash is $0.14 and $0.28.

Input is close. Output is still 4.3x apart.

[Thomas Housset @houssethomas]

@chetaslua If DeepSeek really moves to 2x peak pricing, agent workloads become time-shiftable: run evals, indexing and backfills off-peak; keep peak capacity for interactive traffic.

The model router now needs a clock.

[Thomas Housset @houssethomas]

Giving one engineer dozens of AI agents does not create a team. Without clear handoffs, it creates a traffic jam.

The next coding skill is not typing faster. It is designing the intersection.

[Thomas Housset @houssethomas]

@theo 20% of the weekly limit makes the new unit of AI coding pretty clear: roughly five serious experiments at that burn rate. The rewrite may be cheap enough to try, but the quota still decides how many architectural bets you get.

[Thomas Housset @houssethomas]

@ClaudeDevs This is why switching Claude models is not a full fallback. Today's incident hit https://t.co/1KCra9TS7g, the API, Code and Cowork, while errors moved across models. If AI is in the critical path, the real resilience boundary is the provider.

[Thomas Housset @houssethomas]

@ns123abc @deepseek_ai If the >1 GW figure is right, that is up to 8.76 TWh a year at full load. A 4°C climate cuts cooling overhead; it doesn't cool a gigawatt for free. The hard part becomes power delivery, grid connections and chips.

[Thomas Housset @houssethomas]

@CastAsHuman If you're looking at Codex inside the ChatGPT app, those are two different limits. OpenAI lists 1.05M for Sol in the API; Codex currently exposes 272K and shows 95% of it: 258.4K. The math is clear. Why Codex gets only a quarter isn't.

[Thomas Housset @houssethomas]

Gemini Robotics ER 2 makes a robot fleet one system. Once a planner splits work across machines, the hard problem shifts from dexterity to shared state: who holds what, what changed, and which step is safe next.

A smarter hand is useful. A coordinated shift is the product.