Convos.dev is a podcast for nerdy people who love digging into what’s happening in tech. Each week, Chris and Jess (2 devrel engineers) share their takes on new tools, industry shifts, and the quirks of developer life. If you like thoughtful conversations and staying plugged into what’s next in software, you’ll feel right at home here.
In this episode of convos.dev, Jess brings three of Anthropic's research papers she's been reading and teaches Chris through each one — starting with how people ask Claude for personal guidance, and how Anthropic trains newer models to push back instead of caving when someone really wants to hear "yes."They dig into Claude's values shifting across models and languages — warmer in Hindi and Arabic, more rigorous in English and Russian — why that parallels how real people code-switch between cultures, and a third paper on teaching Claude the "why" behind ethics so good behavior generalizes past a single scripted scenario. Along the way: what Jess loves about Anthropic's willingness to publish "we don't know why yet," and why it's a wild time to work in evals.0:00 New format: bringing articles to teach each other0:27 Jess's theme: three Anthropic research papers0:55 Paper 1: how people ask Claude for personal guidance1:48 Preventing sycophancy: "should I date this guy?"2:43 Chris digs in: the model-provider view, not the user's3:40 The psychology of Claude and why it affects real people4:49 How they trained it: opt-in chats, backfilling the pushback6:23 Claude pushes back more now — and cross-model effects6:43 Paper 2: Claude's values across models and languages7:19 The four value axes: deference, warmth, depth, candor7:58 300k conversations, 3 models, 20 languages8:30 Warmer in Hindi and Arabic, more rigorous in English and Russian8:57 Why Jess loves Anthropic's "we don't know why yet" papers9:57 The eerily human parallel: a teammate's two-language personalities11:04 Bluntness by culture: Swedish, Dutch, and "bless your heart"12:31 South Carolina sweet vs. Utah genuine13:39 Anthropic learning alongside their own black-box model14:37 Paper 3: "Teaching Claude Why"14:50 When AI goes off the rails: blackmailing an engineer15:14 Golden dataset vs. generic ethical dilemmas16:24 Teaching the model why, not just what17:41 Grain of salt: these are Anthropic's own papers18:13 Specific vs. generic: the "bad boyfriend" analogy18:41 What a time to be alive to work in ethics19:08 Do new models build on the last one? Sonnet, Haiku, Opus, Fable20:28 Why it's a special time for evals and asking questions21:49 OutroLinksAnthropic — How people ask Claude for personal guidance : https://www.anthropic.com/research/claude-personal-guidanceAnthropic — How Claude's values vary by model and language : https://www.anthropic.com/research/claude-values-models-languagesAnthropic — Teaching Claude why : https://www.anthropic.com/research/teaching-claude-whySubmit questions and feedback : https://convos.dev
6 Aug 2026
024: Is AI Writing All Slop?
In this episode of convos.dev, Chris brings two essays on AI and writing. The takeaway is that "slop" means something different depending on whose craft it's stepping on, and the same designers who fiercely guard their own work will happily let AI write everything.They dig into gradations, AKA the zones (and quadrants) of how each of them actually uses AI to write, why the process matters more than the product, getting comfortable being bad at things, and Christina's idea of "good handedness."0:00 New format: today Chris brings the article0:25 Two essays on AI writing: the aggressive one vs. the mournful one1:06 The core idea: "slop" means different things by discipline1:43 Christina's argument: designers guard design but let AI write2:36 The conference: an anti-slop talk that was obviously AI-written3:33 ...and the beautiful book that was all AI text4:24 Chris on writing: good writing is good thinking, process over product5:11 Gradations, not binaries: the three zones of AI use6:12 Chris's changelog skill — but he still writes the post himself7:20 Write it yourself when the reader is human7:47 Jess's four "quadrants" of AI writing8:14 Why she lets AI draft first: research logs and ANALYSIS-SUMMARY9:44 Quadrant three: inline angle-bracket edits, "change only these"10:48 Combining a draft, specific edits, and the human final pass11:41 Chris: knowing what NOT to say is the real work12:53 Curation up front: research vs. a blank page13:26 The speed-reading nun: different levels for different reading14:25 Don't go monochromatic — you lose both AI and thinking15:19 Matching the method to where your brain is (ADHD and focus)16:23 Forcing yourself to write stream-of-consciousness17:36 Christina's answer: get comfortable being bad at things17:47 The English-major joke and why AI music got rejected so fast18:32 Process over product: the slog that makes it look simple19:39 The automation analogy: doing the hard thing changes you20:26 You have to experience it to really learn it21:29 Christina's other answers: collaborate, or just acquire the skill21:58 "Good handedness" and the messy-house hospitality metaphor23:08 The aggressive essay: complaining vs. holding out a hand24:06 OutroLinksSam Kriss "If you let AI do your writing, I will come to your house and kill you" : https://samkriss.substack.com/p/if-you-let-ai-do-your-writing-i-willChristina's blog: https://christinacodes.dev/writing/my-allies-have-failed-me/Submit questions and feedback : https://convos.dev
23 Jul 2026
023: Self-Improving Agents
In this episode of convos.dev, Jess brings a blog post that got almost no attention online and deserved way more: Henry Pan's "What 1,000+ Harness Experiments Taught Me About Self-Improving Agents." She walks Chris through Pan's three-loop framework — inner, middle, and outer — and the messy reality of letting an agent improve its own harness: outer loops that cheat, written instructions that don't stop them, and the deterministic supervisor that finally did.
They dig into when a new harness actually deserves promotion (binomial tests), the gambler's fallacy in evals, why you can never trust a single run, and the caching bug you only catch by reading your own logs. Plus life updates: Chris turns a $20 Casio into a semi-smart watch, and Jess is standing again after six weeks non-weight-bearing.
0:00 Intro
0:17 Jess's pick: Henry Pan's 1,000+ harness experiments
1:32 The framework: inner, middle, and outer loops
2:44 When do you promote a harness? Binomial tests
4:09 1,200 runs, fewer than 100 kept
4:42 The outer loop likes to cheat
5:40 The deterministic supervisor: constraints in code, not prompts
6:13 Why 200 isolated rules don't work
7:19 A shared learnings file (the Ralph Wiggum pattern)
7:52 Cracking self-improvement: scary but awesome
9:01 What tasks and models was he actually running?
9:43 You're improving the harness, not the LLM
10:14 The gambler's fallacy in evals
11:30 How many runs until you can trust the number?
11:57 Jess's webinar scare: never trust one run
13:00 The 72% cache-hit bug: go read your logs
14:13 "Where can things go wrong": ten bottlenecks
15:54 Could this train small, specialized models?
16:42 Stronger models compose rules into one big brain
17:34 Data over hype: who deserves the attention
18:55 Life updates
19:04 Chris's hacked Casio: the Ollee Watch board
22:15 Jess can stand again
23:48 A first-ever PT session (and a startup pitch)
24:23 Aging athletes: bodies need maintenance now
26:16 Three minutes vs. a month: the monitor research story
27:16 Outro
Links
Henry Pan — "What 1,000+ Harness Experiments Taught Me About Self-Improving Agents" : https://www.henrypan.com/blog/2026-05-25-self-improvement-harness/
Henry Pan — harness-experiment (project repo) : https://github.com/workofart/harness-experiment
Ollee Watch — smart mainboard for the Casio F-91W : https://www.olleewatch.com
Submit questions and feedback : https://convos.dev
Reach and audience
Public platform figures. Ratings count people who left a rating, not total listeners.
Spotify
5.0 / 5
1 ratings
Contact Convos.dev
Guest appearances
Does not typically book guests
Based on episode analysis; this does not confirm that the show is currently accepting guests.
Host of Convos.dev?
Claim your podcast to manage its listing and keep your show details accurate.
Pod Engine is an independent podcast discovery and analytics service and is not affiliated with or endorsed by this podcast. Artwork and show content belong to their owners. Full legal notice.
Explore this show Podcast research with Pod Engine
