@louisnstudio But isnt every business going to build their own crm along with personal AI assistants like grok bot or dot or muse? Own the data and process.
⚡️ Incredible to see Google pushing the boundaries of AI infrastructure with four distinct TPU systems.
But looking closer to home, Australian developers and businesses are still left waiting. We still don't have NVIDIA L4 GPU support on Cloud Run in local regions like Sydney (australia-southeast1), even though regions like Singapore have it. Without local serverless GPU execution, strict data residency requirements and latency in regulated industries make it impossible for many of us to deploy these AI microservices to Australian enterprise clients. 💼🔒
When can local builders expect parity for Cloud Run GPU workloads down under? ⏳
#GoogleCloud #CloudRun #GCP #Developer #CloudComputing #AIEngineering
Is it just me, or has Claude Code felt noticeably heavier/slower over the last week?
AI engineers, vibe coders, and PMs - are you experiencing the same friction, or did I break my local setup?
#ClaudeCode#opus5.5 #AIEngineering
How do you solve the trust problem with AI-generated code?
AI agents can generate code at a scale too great for humans to review line-by-line. That makes it difficult to ensure the code will do what the developer intended, and it leaves the final output vulnerable to bugs and security risks.
Proof-Driven Development Kit turns plain-language developer intent into a specification in Verus, applying formal methods traditionally used for safety-critical systems. PDD Kit uses mathematical proofs to test agent-generated code against the spec where possible, and clearly indicates where the proofs can’t reach. The result is more reliable agentic code.
Researchers @MatthaiP39719 and @alexlavaee break down Proof-Driven Development in the MSR blog: https://t.co/PoNqwr9EOk
Anthropic just slashed the cost of background subagent status updates in Claude Code 𝗯𝘆 𝟵𝟵%.
If your usage limits are really tight and you want to stretch them as far as they go, the official docs have 𝟯 𝗺𝗼𝗿𝗲 𝘀𝗲𝘁𝘁𝗶𝗻𝗴𝘀 you can change.
→ The built-in Explore uses the same model as your main session by default. Create one named Explore in ~/.claude/agents with model: haiku, and it overrides the built-in one, so every file and code search runs on Haiku
→ Every time one of your own subagents starts, it reads all your CLAUDE.md files, at every level. For the ones that can work from the task description alone, add one line: omitClaudeMd: true
→ A subagent on the same model as your main session follows your /effort if you don't set one for it. Run xhigh, and it runs xhigh too. For simple jobs, give it effort: low
Send this to Claude Code 👇
"Following the official Claude Code subagent docs, help me tune the usage settings of my subagents. Give me a plan first, and don't write anything until I confirm:
1. Create Explore in ~/.claude/agents to override the built-in one: model: haiku, effort: low, omitClaudeMd: true, and only the Read, Grep and Glob tools. Write its prompt around what the built-in Explore is for: read-only, find files, search code, and hand back only its conclusions.
2. Go through my own subagents in ~/.claude/agents and this project's .claude/agents one by one (skip this step if there are none). For the ones that can work from the task description alone, add omitClaudeMd: true. For simple jobs like finding files, running tests or reading logs, set effort: low, and switch to haiku wherever it can do the job.
3. Show me the plan as a table: name, current settings, what you'd change them to, and why.
4. Remind me to run /usage before you change anything and note how much of my usage goes to subagents, then run it again in two days to compare."
Today, we're launching Perplexity Teammates.
They are personal agents for work. And change the paradigm from writing prompts in a session to giving high-level goals. Rolling out to Max users on the web today. Native apps (mobile, desktop) coming soon.
Ontology, Knowledge Graphs, and the Decision Model That Judges Them
Building knowledge graphs with LLMs has a cost structure that is easy to misread. Extraction is cheap and getting cheaper. Verification is where the money goes.
For every candidate entity and every candidate relation, something has to decide whether it is true to the document. That judgment repeats tens of thousands of times, which is the job a decision model exists for.
The model in question is Jev, a small decision model from TypeSafe built to take a claim, weigh the evidence and commit to an answer.
The pattern it challenges:
- One large model handling extraction, summarisation and community building, with every step billed at frontier rates
-The model that proposes a fact also being the one that grades it
-Reasoning models adding a deliberation meter, paying for pages of thinking to collect one-word yes/no verdicts
OntoSmith splits the work into three layers:
-A cheap mid-tier model proposes triples (Claude Haiku 5.5 in the reference setup). It reads and suggests. It does not decide.
-A fitted pair call and then Jev judge every candidate before anything reaches the graph
-A frontier model is reserved for the genuinely hard calls
The shared map is the ontology: versioned, hash-pinned packs for open-domain text, legal contracts and biomedical literature over a domain-neutral core. For a document type with no ready ontology, one is drafted for approval. The graph lives in Neo4j, one graph per ontology build.
This is not RAG versus Graph RAG. Search handles answers that sit in one paragraph. Graphs add value when relations span pages: +0.080 recall@20 (0.631 to 0.712) on a pinned 116-question set over long contracts, and exactly zero on short abstracts.
The measured figures:
-3.8 cents per document across 500 documents in the flagship build
-About 7 cents per document projected for judging alone in the default configuration
-Corrected graph precision of 0.911 [0.892 to 0.928], revised down after a re-audit by a judge from a different model family, with the working shipped alongside
One lesson from the build: a decision model on default parameters is someone else’s operating point. Thresholds were fitted stage by stage, and experiments that missed their pre-registered bars stayed in the record as failures.
OntoSmith is MIT-licensed, on PyPI (pip install ontosmith) and on GitHub. 🔎
By Dongbum Lee
Github https://t.co/rLG0vHYWVZ
Article https://t.co/sGbdTp9wR7
#OntoSmith #Neo4j #DecisionModels #RAG #LLM
--
Connected Data London 2026 | 11–12 November | Leonardo Royal Hotel London Tower Bridge
🎟 Tickets on sale now - 🔔 Just a few days left!
https://t.co/xCS7dkglKm
📺 Sponsorship opportunities available. Contact [email protected] for details.
#KnowledgeGraph #GraphRAG #Ontology #Graph #AI #DataScience #GraphDB #SemTech
We will all be learning in coming weeks & months, that most human-made software:
▫️ Was really slow
▫️ Had many security vulnerabilities
▫️ Was under-specified and incorrect
AI will rewrite it all. Many times. It will make it fast, cost-efficient, and secure. It will think about failure modes we couldn't wrap our heads around.
It will prove, formally verify, load test, harden. It will turn memory-unsafe foundations into robust, optimized ones. If you ask, it'll thoroughly document it, teach you about it, make diagrams, animations, and slide decks.
In many ways, we are witnessing the true birth of software engineering.
Memory in AI Agent Harness.
LLMs are stateless, so every bit of memory in an AI agent is something the harness saves to disk and loads back in when it needs it.
most good agent harnesses split that memory into a few layers, let's look at each one of them:
• context: what the model can see in the current session. when it gets close to full, the harness summarizes the older messages to make room.
• notes: a progress file the agent keeps updating while it works, so if a session ends halfway the next one can read it and continue from there.
• semantic: facts about you, your team and your project, like your preferences or which repo you work in. Hermes keeps these in MEMORY.md and USER.md.
• procedural: how to do a specific job. when the agent figures out something hard, it writes it down as a skill and uses that skill the next time.
• episodic: a log of every past session that the agent can search when it needs to remember what happened before.
First two only last as long as the session or the task. Other three are persistent, so they're still there the next time you start the agent.
save this for the next time someone asks how agents remember.
follow @Saboo_Shubham_ for more on Agent Harness.