AI News Briefing

Scope

What came from Horizon, and what I added

Horizon is a general-purpose aggregator. Out of the box it's set up for general tech and finance news, and on this machine it didn't produce a briefing at all until two bugs were fixed.

From Horizon (open source)

The pipeline engine and its collectors for RSS, Hacker News, Reddit and GitHub
The LLM scoring framework and profile system
Duplicate merging and quota-based selection
Write-ups using web search and past briefings
Markdown archive and email delivery

What I added

Tool selection: compared five approaches against written requirements
AI-only scoring profile: relevance rule and rubric (profiles/ai-news)
Sources and quotas: 16 AI sources, with combined Reddit feeds to get around blocking
Two bug fixes to Horizon's code: API error handling and rate-limit retry
Tuning: thresholds, model choice, and concurrency under rate limits
Automation: scheduled task, logging wrapper, setup scripts
Documentation: this case study and an operator guide

The goal

One short, trustworthy AI briefing a day

AI news is spread across lab blogs, tech press, newsletters, Reddit, Hacker News and GitHub, and most of it is noise. I wanted a "what's new" email each morning that only covered things that mattered, explained why they mattered, and didn't need me to keep anything running.

My requirements were:

Pull from the primary sources (the labs themselves) as well as the press and the communities discussing them.
Filter by importance, not popularity, and drop anything that isn't really about AI.
Merge the five articles about the same launch into one entry.
Add enough background that a headline makes sense on its own.
Run unattended every morning and arrive by email.

Research

Build, borrow or buy

Before building anything I surveyed what already existed on GitHub, in the n8n template library and in community write-ups. I compared five approaches against the requirements.

Option	What it is	Strengths	Why not / why	Verdict
Horizon Thysrael/Horizon · ~12.8k ★ · MIT	Open-source Python news aggregator with LLM scoring	Many source types, per-profile scoring rubrics, AI-based duplicate merging, background enrichment with web search, email delivery	Covered every requirement and is actively maintained. Needed AI-specific configuration and two bug fixes	Chosen
agents-radar ~1.2k ★	A ready-made daily digest published by someone else	No setup: subscribe by RSS or Telegram	Fixed sources, focused on AI developer tools, no control over filtering	Not chosen
n8n templates	Visual workflows: RSS → LLM summary → email	Easy to read and change	Needs an n8n server. Templates do simple keyword filters, with no scoring rubric or duplicate merging	Not chosen
Claude scheduled task	An AI assistant prompted to search the web each morning	Almost no setup and no API key	Results vary from day to day, with no fixed sources or repeatable scoring	Not chosen
Custom script	Writing my own fetch → summarise → email script	Total control	Would rebuild most of what Horizon already does well	Not chosen

Architecture

How it fits together

Collection, filtering and scheduling run locally. Only the LLM calls, web search and email leave the machine.

Orange outlines mark what this project added; grey is Horizon or third-party services. Horizon's eight steps run in a fixed order. Scoring sends one LLM request per item and returns structured JSON (score, reason, summary, tags). Write-ups only happen after quotas are applied, so no tokens are spent on stories that won't make the briefing.

How a hundred stories become ten

Real counts from a weekday test run, drawn to scale. The score threshold does most of the work.

Making it work

Four problems found in testing, and how they were fixed

The software installed cleanly but didn't work as intended on the first real runs. Each problem was spotted in the run logs or the briefing itself, reproduced in isolation where possible, and fixed with the smallest change that solved it.

1. Every LLM call failed

Symptom

All 176 scoring requests to Claude came back as 400 Bad Request, so the briefing would have been empty.

Diagnosis

A minimal API call with the same key worked. Adding settings one at a time isolated temperature: the API now rejects it with "Extra inputs are not permitted". Horizon has a fallback that retries without temperature, but it only recognised older wordings of that error.

Fix

One line added to the error-matching function, so the existing fallback runs.

Result

Every call succeeds on the first fallback. Failed requests aren't billed, so the broken run cost nothing.

src/ai/client.py · _is_temperature_unsupported()         or "only 1 is allowed" in lowered
+        or "not permitted" in lowered  # Anthropic: "temperature: Extra inputs are not permitted"
     )

2. Reddit was blocking the scraper

Symptom

Collection took about 13 minutes, and Reddit posts arrived without their comments.

Diagnosis

Reddit returns 403 to automated requests for its normal pages and its JSON API. Horizon tried both routes for every subreddit and every post, about 20 seconds per blocked attempt, before falling back to RSS. The project's issue tracker showed other users hitting the same thing. RSS was rate-limited too, returning 429 for about half of rapid requests.

Fix

Read RSS directly and combine the five subreddits into two multi-subreddit feeds, so there are 2 requests a day. Add retry with linear backoff on 429.

Result

Reddit now takes 2 requests a day instead of dozens of blocked ones, and the whole run finishes in about 6 minutes. In testing, a feed refused three times succeeded on the fourth attempt.

src/scrapers/rss.py · _fetch_feed()+        # Retry rate-limited feeds (Reddit RSS often answers 429) with backoff.
+        for attempt in range(4):
+            response = await self.client.get(feed_url, follow_redirects=True,
+                                             headers={"User-Agent": RSS_USER_AGENT})
+            if response.status_code != 429 or attempt == 3:
+                break
+            await asyncio.sleep(15 * (attempt + 1))
         response.raise_for_status()

3. Half the briefing wasn't about AI

Symptom

The first working briefing had 4 items. Two were general tech stories (an SSH vulnerability, a desktop-app rewrite) that came in through Hacker News.

Diagnosis

The built-in scoring rubric is for all of technology, so an important non-AI story scores as highly as an AI one. The threshold of 7 was also too strict for a niche topic.

Fix

A custom ai-news profile with its own matching and scoring prompts and a relevance rule: non-AI items score at most 3, and items that mention AI only in passing score at most 4. The rubric bands were rewritten for an AI reader, and the threshold was lowered to 6.

Result

On the same day's sources: 10 items, all about AI.

Before · general tech rubric

4 items, 2 off-topic

After · ai-news rubric

10 items, 0 off-topic

4. Write-ups silently disappeared

Symptom

One run produced a briefing where all 10 items had only a one-line summary. The background and impact sections were missing, though the run reported success.

Diagnosis

The enrichment step failed straight after a burst of about 100 scoring requests at concurrency 4. A short rerun with fewer items enriched every item correctly. That points to request-rate limits on a new API account rather than a code fault.

Fix

Reduced concurrency to 2 for scoring and 1 for write-ups. That trades about a minute of run time for headroom under the rate limit.

Result

The first scheduled run after the change wrote up every item in full.

Key decisions

The trade-offs behind the design

Horizon

over n8n templates or an AI scheduled task

A fixed set of sources with a written scoring rubric gives a briefing I can tune and trust from day to day. A prompt-only approach would produce something different every morning.

Combined Reddit RSS feeds

over Reddit's scraper and API routes

Reddit blocks the other routes, so the only reliable path is RSS. Two combined feeds keep the request count low enough to stay under its rate limit, at the cost of comment threads.

Claude Sonnet 5.5

over the cheaper Haiku 5.5

Scoring and write-ups need judgement, and quality is the point of the briefing. Haiku stays available as a cost lever with a one-line config change, once a few weeks of usage data show what a run actually costs.

Python virtual environment

over Docker

Horizon ships with Docker support. For one daily job on one Windows machine, Docker Desktop would add WSL2, a background service and memory overhead, plus one more thing that has to be running at 6:30, with no benefit at this scale. Docker becomes the right choice if the job moves to a server so it runs while the PC is off.

Results

What it produces

~10items on a typical weekday, from about 100 collected

0off-topic items in the 3 runs since the custom rubric

6 min 26 sfull run, measured through the scheduler

160–200kLLM tokens per run, about 85% input

The Ultimate MoE Training Handbook: training Mixture-of-Experts models with Megatron-LM

★ 6.0/10 · Reddit: LocalLLaMA + MachineLearning

Summary

A new handbook for engineers who have trained dense models explains how to train Mixture-of-Experts models with Megatron-LM, from a hardware test and dense baseline to a first single-node MoE run. The authors say it draws on about 800k GH200 GPU-hours on the Alps cluster at CSCS.

Background

An MoE model replaces each dense feed-forward layer with many smaller "expert" networks and a router that sends each token to only a few of them, cutting compute per token but adding problems such as balancing load across experts.

Impact

Teams moving from dense to MoE training get a consolidated starting point: a baseline comparison, recommended flags and scripts. Teams on other hardware should validate the settings themselves.

#Mixture-of-Experts#Megatron-LM#LLM training#handbook

One item from a real briefing, shortened. The model was told to separate vendor claims from verified results, and the write-up does that: "the authors say", and a note that the recommendations aren't benchmarked.

Skills and tools

What this project involved

Evaluating tools

Writing requirements first, surveying GitHub, n8n and hosted options against them, and justifying the choice and its trade-offs.

Configuring LLMs

Writing a scoring rubric and relevance rules as prompts, choosing the model, and tuning thresholds, token use and concurrency under rate limits.

Fixing someone else's code

Reading logs, isolating failures with minimal reproductions, interpreting HTTP 400, 403 and 429 responses, and making small, targeted patches to an unfamiliar codebase.

Automation and ops

Python virtual environments, PowerShell, Windows Task Scheduler, SMTP over SSL, and keeping secrets out of version control.

How it was built. I built this with Claude Code, Anthropic's AI coding agent, working as a pair. I set the goal and requirements, chose between the options, made the trade-off decisions above, and reviewed each test briefing. Claude Code did the source research, wrote the configuration and custom profile, read the logs, and made the code fixes. Every change was checked against a real run before moving on.

Next steps

Where it could go

Run it in a container on a small cloud server, so it no longer depends on the PC being on at 6:30. This is where the Docker support starts to pay off.
Both fixes have been submitted to Horizon as pull requests, since other users hit the same errors: #220, the temperature error and #221, the rate-limit retry. Both include tests.
Score with a cheaper model and write up with a stronger one, once there's enough usage data to measure the cost and quality difference.
Add a feedback loop: mark items as useful or not, and feed that back into the rubric.