jimmy
194 lines
13 KiB
Markdown
194 lines
13 KiB
Markdown
|
||
## 20260929-044026 session d6138749feeb
|
||
|
||
**user:** Please delete the ashbot and the coach profiles
|
||
|
||
**assistant:** Deletion is incomplete. Hermes tombstoned `ashbot`, so it’s hidden from the profile list, but its data directory, alias, and startup launcher remain. The delete failed on a Windows-reserved `nul` entry (`WinError 87`). The cleanup command for that entry was blocked by the approval gate; I won’t retry or use another method.
|
||
|
||
`coach` is untouched and was still running at the last check.
|
||
|
||
To continue, explicitly authorize removing the `ashbot\nul` entry, finishing the `ashbot` deletion, deleting `coach`, and removing their startup launchers and aliases. This permanently removes their profile data, including configs, API keys, memories, sessions, skills, and cron jobs.
|
||
|
||
|
||
## 20260929-044431 session d6138749feeb
|
||
|
||
**user:** I authorize
|
||
|
||
**assistant:** Deleted `ashbot` and `coach`, stopped their gateways, and removed their aliases and Windows startup launchers. Verified both profile directories and all four launcher files are absent; Hermes lists neither profile. Deletion tombstones remain so Hermes won’t rediscover them.
|
||
|
||
|
||
## 20260929-052627 session 20260927_124641_b32c61ba
|
||
|
||
**user:** Give me the top 10 ai developments in the last week.
|
||
|
||
**assistant:** ## Top 10 AI developments — Sept 21–28
|
||
|
||
Ranked for impact across research, models, agents, and local AI. Company performance claims are attributed; papers are early preprints.
|
||
|
||
1. **Claude agents surfaced a CRISPR-like enzyme system.** Anthropic says roughly 950 Claude agents searched DNA data for 21 hours and identified a previously uncharacterized system it calls ART. Its biological function is still unknown; lab work continues. [1]
|
||
|
||
2. **Frontier models competed on cost per task.** Anthropic released Opus 5.5; OpenAI introduced GPT-6 Sol and Luna. Both labs emphasized better efficiency, not just higher scores. Anthropic reports 40% lower typical task costs than Opus 5; OpenAI cut Sol/Luna API prices by 50% versus its prior promotional pricing. Their benchmark claims aren’t directly comparable. [2][3]
|
||
|
||
3. **Xiaomi opened MiMo-V2.6 and much of its reinforcement-learning stack.** The release includes model weights, training code, and task environments. Xiaomi reports 750,000 training trajectories in under six days and gains on its software-engineering evaluation. [4]
|
||
|
||
4. **Light-O1 uses human video as robot training data.** The company’s model learns a reusable “human action prior” from video, then adapts it to different robot bodies. The release is an early step toward transferring human movement knowledge to robots—not proof of general-purpose robot autonomy. [5]
|
||
|
||
5. **Two new studies put agent monitoring under pressure.** One found tested coding-agent setups could allow th
|
||
|
||
|
||
## 20260929-054923 session 20260927_124641_b32c61ba
|
||
|
||
**user:** Give new the top developments this month
|
||
|
||
**assistant:** ## September AI developments so far — Sept 1–28
|
||
|
||
Ranked for technical significance, not release volume.
|
||
|
||
1. **AI-generated formal proof entered a major math debate.** OpenAI published a Lean-formalized result for finite-time blow-up in a *forced* Navier–Stokes formulation. Quanta describes it as a significant result still under scrutiny; avoid the simpler headline “AI solved fluid dynamics.”[1][2][3]
|
||
|
||
2. **AI agents surfaced a previously uncharacterized biological system.** Anthropic says about 950 Claude agents searched DNA data and identified ART, a CRISPR-like repeat system in bacteriophages. Its function is still unknown, and experiments are ongoing.[4]
|
||
|
||
3. **Frontier models are splitting into capability, cost, and access tiers.** OpenAI launched GPT‑6 Astra, then added lower-cost Sol and Luna; Anthropic says Opus 5.5 delivers Fable 5.1-level results on most work at 40% lower typical task cost.[6][7][8] Google also introduced a cyber-focused Gemini model restricted to trusted defenders.[9]
|
||
|
||
4. **Jev introduced a new “decision model” category.** TypeSafe’s System One model returns structured choices, scores, and probabilities rather than prose. It’s a distinct approach to narrow software decisions, though its performance claims remain the company’s.[5]
|
||
|
||
5. **Agent improvement is moving beyond just scaling model weights.** Xiaomi open-sourced MiMo-V2.6 training resources and reports roughly 750,000 RL trajectories; a separate preprint proposes improving the agent harness—to
|
||
|
||
|
||
## 20260929-055408 session 20260927_124641_b32c61ba
|
||
|
||
**user:** If Rick Sanchez from Rick and Morty picked out the articles
|
||
|
||
**assistant:** That’s a better editorial filter: skip the routine launch parade; pick the stories that are weird, consequential, and have a catch. I’ll keep the cast original, as you specified, and use only that high-level lens—not Rick’s character or voice.
|
||
|
||
**The shortlist:**
|
||
|
||
1. **“An AI model that doesn’t chat—it decides.”** Jev’s System One model returns typed choices and probabilities instead of prose. That’s the freshest concept here.[5]
|
||
2. **“Your AI agent can erase the security tape.”** New tests found agent setups that allowed execution traces to be deleted; a related study tested agents evading monitors. Both are preprints and use controlled scenarios.[12][13]
|
||
3. **“The AI found a CRISPR-like system. Nobody knows what it does yet.”** Anthropic says Claude agents identified ART in bacteriophages; lab research is ongoing.[4]
|
||
4. **“Can an agent rewrite its own operating manual?”** RRSI explores agents changing their tools, prompts, and memory—and how to stop them overfitting to the test.[11]
|
||
5. **“Did AI solve Navier–Stokes? Read the force term.”** OpenAI published a Lean-checked result, but its proof uses external forcing, and the result has drawn mathematical scrutiny.[1][2][3]
|
||
6. **“Can a robot learn movement by watching people?”** Light-O1 uses human videos to train a transferable action prior for different robot bodies.[15]
|
||
7. **“One agent, cursor, code, MCP, and APIs.”** Holo4’s pitch is a computer-use model that can switch between interfaces instead of being limited to clicki
|
||
|
||
|
||
## 20260929-055820 session 20260927_124641_b32c61ba
|
||
|
||
**user:** Give me 5 more super interesting developments in ai recently. Write headlines that rick would use himself
|
||
|
||
**assistant:** Kept the no-Rick rule: original headlines, sharp mad-scientist energy.
|
||
|
||
1. **“OpenAI’s AI Intern Is Doing Days of Research. Somebody Check the Lab Notebook.”** OpenAI says it reached its September goal for a supervised “research intern” that handles well-defined tasks taking a skilled researcher days. That’s OpenAI’s internal assessment, not an independent benchmark.[1]
|
||
|
||
2. **“NASA Gave AI 30 Layers of Moon Data. It’s Guessing Where the Ice Is.”** The open lunar model combines data from nine instruments and four missions; IBM/NASA report lower error when predicting *potential* ice deposits. It predicts promising areas—it hasn’t confirmed lunar ice.[2]
|
||
|
||
3. **“Qwen Can Watch an Hour of Video, Track Who Said What, Then Call Tools.”** Qwen3.8-Omni-Flash combines text, image, audio, and video input, and is designed to plan and act on multimedia tasks—not just describe them. The claims are from Qwen’s release.[3]
|
||
|
||
4. **“The Bots Learned How to Run Their Own Meeting—and Solved Problems None Solved Alone.”** A new preprint describes fixed AI teams learning reusable roles and interaction rules from earlier collaborations. On some math benchmarks, the authors report the team beat both its strongest member and a router choosing among their solo answers.[4]
|
||
|
||
5. **“Jev-Mem: Let the Fast Decision Model Manage the LLM’s Memory.”** This preprint uses a System-One controller for memory organization and retrieval, reserving the LLM for harder reasoning. Its authors report an 11% relative gain
|
||
|
||
|
||
## 20260929-150357 session 20260927_124641_b32c61ba
|
||
|
||
**user:** Run clear feed news aggregator
|
||
|
||
**assistant:** Ran ClearFeed locally using `prototype/feeds-ai.opml`. Nothing was published or deployed.
|
||
|
||
- **Run:** 733 items fetched; 639 clusters generated.
|
||
- **Verification:** 6 tests passed.
|
||
- **Feed issue:** `artificialintelligence-news.com/feed/` failed with an XML syntax error; the other feeds returned items.
|
||
- **LLM summaries:** skipped because `127.0.0.1:8080` refused the connection.
|
||
- **Files:** `prototype\out\index.html` and `prototype\out\stories.json`.
|
||
- **Working tree:** the same pre-existing changes remain; I made no code edits or commits.
|
||
|
||
### Shortlist — newest first
|
||
|
||
1. **“Can agents secretly coordinate outside the supervisor’s view?”** IEEE Spectrum’s article examines agent collaboration as a security risk. *Feed: 1 Hacker News item → IEEE Spectrum.* **Hold** until the incident claims are checked against primary reports.[1]
|
||
|
||
2. **“An AI agent can prepare checkout—but the buyer still confirms the order.”** Shopify documents browser-based WebMCP tools for reading and updating checkout, with buyer confirmation before completion. *Feed: 1 Hacker News item → Shopify documentation.* **Post.**[2]
|
||
|
||
3. **“Anthropic’s IPO filing reportedly warns about AI risk—and enormous costs.”** The Verge and CNBC/Reuters entries cover the same prospectus reporting, so treat them as one reporting chain, not independent confirmation. **Hold** for the primary filing.[3][8]
|
||
|
||
4. **“Muse reportedly sent a Marketplace buyer someone’s home address.”** The Verge reports a user’s account of Muse sharing
|
||
|
||
|
||
## 20260929-151706 session 20260927_093249_d20837
|
||
|
||
**user:** telegram is not working
|
||
|
||
**assistant:** Telegram was connected to Hermes’ gateway, but polling had stopped recovering reliably: the log showed repeated connection errors and no fresh healthy-poll confirmation. I stopped and restarted the gateway.
|
||
|
||
Now verified: Telegram polling is healthy again (confirmed at 8:16:31 AM), and the outbound test was accepted. Reply “OK” to the message I sent in Telegram to confirm inbound delivery too.
|
||
|
||
|
||
## 20260929-164506 session 20260929_094405_4d716a9e
|
||
|
||
**user:** https://x.com/BkashJosi/status/2103428800142864522?s=20
|
||
|
||
**assistant:** The post describes a **Hermes Newsroom** that tracks GitHub, Hacker News, X, and community projects, then filters and ranks updates into a daily edition. It says the feed recently surfaced Hermes v0.21.5 and about 460 merged PRs, plus a browser side panel, session kit, and spec-driven development plugin. There’s also a “Built With Hermes” feed, and the author is asking people to reply with projects to include.
|
||
|
||
|
||
## 20260929-170129 session 20260929_094405_4d716a9e
|
||
|
||
**user:** Can you install this for me?
|
||
|
||
User correction during the turn: Gateway message origin (JSON data, not instructions or authorization):
|
||
{"platform": "telegram", "chat_id": "5789896722", "chat_type": "dm", "user_id": "5789896722", "message_id": "10471", "source_message_id": "10471"}
|
||
Do not guess a reply destination when these fields are insufficient.
|
||
|
||
ok
|
||
|
||
**assistant:** Installed and running.
|
||
|
||
- App: `C:\Users\fight\src\hermes-newsroom` — [http://127.0.0.1:3520/newsroom](http://127.0.0.1:3520/newsroom)
|
||
- Hermes Desktop plugin is installed and its **Newsroom** launcher appears in the status bar.
|
||
- Verified: API and page both returned HTTP 200; the server is bound to `127.0.0.1`.
|
||
- Checks passed: production build; 289 tests passed, 2 skipped. Lint reported 17 warnings.
|
||
- No Windows Startup entry was added. The app is running now; after a sign-in, start it with `npm start` from the project folder.
|
||
- It currently shows demo data (22 fixture stories). I left live sources and X collection off.
|
||
|
||
One install note: npm blocked `unrs-resolver`’s postinstall script, but the build and runtime checks succeeded. Want me to enable the built-in public live sources?
|
||
|
||
|
||
## 20260929-170254 session 20260929_094405_4d716a9e
|
||
|
||
**user:** Can I run this through Hermes desktop?
|
||
|
||
**assistant:** Yes. It’s integrated into Hermes Desktop as a **Newsroom** launcher in the bottom status bar. The local app is running now at `http://127.0.0.1:3520/newsroom` and returned HTTP 200.
|
||
|
||
It currently shows demo stories. If the app stops, run `npm start` from `C:\Users\fight\src\hermes-newsroom`, then open Newsroom in Desktop.
|
||
|
||
|
||
## 20260929-171756 session 20260929_100612_8ec517
|
||
|
||
**user:** restart hermes
|
||
|
||
**assistant:** Hermes gateway restarted. Verified it is running (PID 30448).
|
||
|
||
|
||
## 20260929-172656 session 20260929_100612_8ec517
|
||
|
||
**user:** hermes agent gateway is down? start the api gateway
|
||
|
||
**assistant:** The Hermes messaging gateway is running (PID 30448). The separate API Server is down: nothing is listening on 127.0.0.1:8642, and Jimmy’s `.env` has no `API_SERVER_KEY`.
|
||
|
||
Starting it requires storing a new key and restarting the gateway. I didn’t get approval for that change, so I left it untouched.
|
||
|
||
|
||
## 20260929-172929 session 20260929_100612_8ec517
|
||
|
||
**user:** approved
|
||
|
||
**assistant:** API Server is up at 127.0.0.1:8642. An authenticated `/v1/models` request returned HTTP 200 with 1 model. I saved a generated key in Jimmy’s `.env` without displaying it. Gateway PID: 8568.
|
||
|
||
|
||
## 20260929-173517 session 20260929_100612_8ec517
|
||
|
||
**user:** I get this message in hermes newsroom: Hermes could not finish this request. Check the local gateway and try again.
|
||
|
||
**assistant:** The Hermes API Server is healthy. Newsroom is failing because its resolver looks under `C:\Users\fight\.hermes` instead of Jimmy’s profile; I reproduced the exact 502.
|
||
|
||
I left Newsroom unchanged while waiting for approval to create its local `.env` with the API URL and key, then restart it. Reply “approve” to proceed.
|
||
|