This Week in All Things AI - Week 19-2026
This newsletter exists to give busy, technically curious people a curated weekly snapshot of how AI tools, services, and use cases are evolving in practice, drawn directly from the All Things AI Telegram group.
This Week in All Things AI (Week 19, 2026) is my attempt to distill a very full seven days in the Telegram group into something you can skim in a few minutes and then dive into at your own pace. From real‑time voice agents and Harvard’s study of OpenAI’s o1 beating human doctors on complex triage reasoning, to finance workflows, frontier‑scale models, and the evolving AI infrastructure stack, this week really shows how fast ideas are turning into production‑grade tools and services.
As you read through the day‑by‑day notes below, you’ll see threads that keep resurfacing: security and agent design in the wake of a Grok prompt‑injection exploit, attempts to tame AI costs like Alibaba Cloud’s MaaS token plan, and new model bets such as SubQ’s 12M‑token SSA architecture. You’ll also find Anthropic and Perplexity deepening their focus on finance and health, Anthropic’s unexpected compute partnership with SpaceX and higher Claude limits, plus community debates on Chinese models vs Claude, programs for Indian student builders, and some very nerdy but important meta‑topics like HTML‑first formats and how to think about agent skills.
Readers who think others in their family, friends and acquaintances who are curious in knowing more about rapidly evolving AI tools/services/use cases and would benefit from being subscribers of this weekly newsletter are encouraged to share this publication link to them and invite them to subscribe.
The following messages were posted on the ' All Things AI ' Telegram group from Sunday 3rd May 2026 to Saturday 9th May 2026
Sunday 3rd May 2026
via Mahimai Raja
A curated, developer friendly learning path for building real-time voice AI agents from your first STT call to scaling production telephony.
I built this because Voice AI is moving fast and I couldn't find a single place that walks a developer from "what is a voice agent" through to production telephony, evals, and the FCC/EU AI Act stuff you actually need to know before shipping.
Every citations are verified and active, tagged Beginner/Intermediate/ Advanced, and grouped so you can read it top-to-bottom:
GitHub - mahimairaja/voiceai: Set of 📝 with 🔗 to help those building Voice AI agents 🎙️🤖
Harvard study: OpenAI's o1 correctly diagnosed 67% of emergency room patients using electronic records and a few sentences from nurses, vs. to 50-55% for triage doctors.
A groundbreaking Harvard study has found that AI systems outperformed human doctors in high-pressure emergency medicine triage, diagnosing more accurately in the potentially life and death moments when people are first rushed to hospital.
Performance of a large language model on the reasoning tasks of a physician
New AI Model Beats Doctors at Clinical Reasoning, Diagnosis
AI outperforms doctors in Harvard trial of emergency triage diagnoses
Nice X thread by the author of the study
1/ Our new study on AI and physician reasoning just came out in @ScienceMagazine. As co-senior author, I'm excited about our findings, and I do think AI will reshape medicine. But after seeing some of the discussions, I'm also worried about how our findings may be.
Monday 4th May 2026
Most of you know that I'm a big fan of Wispr Flow and mention it whenever I tell people about voice dictation.
Tanay Kothari, CEO of Wispr Flow on CNBC's Young Turk's with Shereen Bhan. It's around an hour long.
Tuesday 5th May 2026
via Alex
Grok morse code prompt injection led to transfer of 200k$ in DRB token.
That is why we need proper frameworks for agents identity, secrets & session management with cryptographic intent verifications.
saw the thread. here's what I can piece together from the tweet + replies:
someone (ilhamrfliansyh) pulled an exploit on a $DRB token interaction through bankrbot today
the play involved some kind of prompt injection / array manipulation — one reply mentions an array named
via Alba Chung of Alibaba Cloud.
Alibaba Cloud MaaS Token plan released! It uses Credits as a unified billing unit, supports text and image generation models, and is compatible with popular AI programming tools and agents. It offers stable performance and enterprise-grade data security.
Perfect for enterprises start want to adopt to AI tools with controlled budget and price performance.
Feel free to PM me for more details.🙌🏼
Insert your favourite variant of the 'Shut up and Take My Money' meme.
via Alexander Whedon.
Introducing SubQ - a major breakthrough in LLM intelligence.
It is the first model built on a fully sub-quadratic sparse-attention architecture (SSA),
And the first frontier model with a 12 million token context window which is:
- 52x faster than FlashAttention at 1MM tokens
- Less than 5% the cost of Opus
SubQ is available for early access today, alongside our coding agent, SubQ Code.
Wednesday 6th May 2026
Anthropic Launches Claude AI Agents for Finance Workflows.
Anthropic unveiled 10 Claude AI agent templates designed for banks, insurers, and financial firms, tackling tasks like drafting pitchbooks, reviewing earnings transcripts, reconciling ledgers, and screening KYC files. Powered by Claude Opus 4.7 and connected to data sources such as FactSet, S&P Capital IQ, and Moody's, the agents integrate with Microsoft 365 apps and output source-attributed results for compliance.
New for financial services: ready-to-run Claude agent templates for building pitches, conducting valuation reviews, closing the books at month-end, and more.
Perplexity launching Perplexity Computer for Professional Finance. Finance teams can bring licensed data from providers like Morningstar, PitchBook, Daloopa, and Carbon Arc into Computer.
We’ve also added 35 dedicated finance workflows for the work analysts repeat every week.
Every output is traceable. Computer shows where the data came from and links directly to the source behind key numbers. Click any citation or hyperlinked number to open the underlying SEC filing, earnings transcript, market data page, or licensed source.
Perplexity and Computer now connect to premium health sources, starting with NEJM and BMJ Group, with 9 more medical journals and clinical databases on the way.
Thursday 7th May 2026
This was not on my bingo card 🤯
Anthropic agrees to partnership with SpaceX to use all compute capacity at Colossus 1 data center.
The partnership enables Anthropic to double Claude Code rate limits for Pro, Max, and Team plans, remove peak hours limit reductions for Pro and Max, and substantially increase API rate limits for Opus models.
Higher usage limits for Claude and a compute deal with SpaceX
Friday 8th May 2026
Saturday 9th May 2026
via Benjamin
Great post from Thariq at Claude about HTML > Markdown files.
Perplexity published their internal manual for building agent skills and they state that skills require a new way of thinking for developers.
Designing, Refining, and Maintaining Agent Skills at Perplexity
Antirez Releases ds4.c, a Native Inference Engine for Running DeepSeek V4 Flash on 128GB MacBook Pro. Antirez, the Redis founder, launched ds4.c, enabling local inference of the 284B MoE DeepSeek V4 Flash model on high-end Macs. The project uses selective quantization, Metal execution, and a 1M token context window with disk-backed KV cache. It supports coding agents and advances architecture-aware local inference for frontier-scale open models.