Guru99 AI Report › News Letter › Current Edition

1M token context at 1/7th the price – inside DeepSeek V4

ALSO: AI beats Harvard’s ER doctors, DeepL cuts 250 jobs

Krishna Rungta May 15, 2026

Welcome to Guru99 AI Report!

Top Story: Welcome back! Buckle up — DeepSeek V4 is rewriting the pricing rulebook, AI just out-diagnosed Harvard physicians, and a single hour of scanning exposed a Linux flaw lurking since 2017. Five stories you’ll want to read before tomorrow’s headlines catch up.

🚀 DeepSeek V4 Just Shook the AI World — Again!

Brief Buzz: Chinese AI lab DeepSeek is back in the spotlight with previews of its new flagship models, DeepSeek-V4-Pro and DeepSeek-V4-Flash. The release doubles down on what made the lab famous earlier: serious performance at a fraction of the cost of Western frontier labs.

V4-Pro packs 1.6 trillion total parameters (49B active) and leads all open models in reasoning, math, STEM, and coding — competing with GPT-5.4-High and Claude-Opus-4.6-Max.
V4-Flash is the lightweight sibling at 284B/13B active parameters, built for faster, cheaper deployment.
A new hybrid attention system (Compressed Sparse Attention + Heavily Compressed Attention) slashes compute costs while expanding context to 1 million tokens.
Pricing? Just $3.48 per million tokens versus $25 for Claude Opus 4.6 and $15 for GPT-5.4.
Available now via the DeepSeek website (Expert Mode/Instant Mode) and API.

💡 Why Should You Care?

Cheaper, more capable AI means lower costs eventually trickle down to the apps and services you rely on. DeepSeek’s aggressive pricing could force US labs to cut their rates, while its million-token context window unlocks smarter agents that can remember and reason over much more of your data in a single session.

📰 DeepL Cuts 25% of Staff in Sweeping AI Restructuring

Brief Buzz: German translation startup DeepL is cutting roughly 250 jobs — about 25% of its 1,000-strong workforce, with CEO Jarek Kutylowski citing a “massive structural shift” driven by AI. The Cologne-based Google Translate rival joins a growing list of tech firms restructuring in the AI era.

CEO says the move means “fewer layers, faster decisions” and less back-and-forth slowing teams down, as the company adapts to work effectively with AI.
DeepL now competes head-on with ChatGPT and other AI tools offering free translation, squeezing its core business.
The cuts mirror a wider trend: Meta announced 10% layoffs last month, while Microsoft is offering buyouts to 7% of its US staff.
DeepL raised $300 million at a $2 billion valuation in 2024 and has been weighing a US IPO.

💡 Why Should You Care?

This isn’t just another tech layoff — it’s a signal that AI is now eating its own. A company built on AI translation is being disrupted by even more powerful generalist AI tools. For workers everywhere, it’s a reminder that no industry, even AI itself, is immune to automation’s ripple effects. For users, expect translation to keep getting cheaper and more bundled into everyday apps like ChatGPT.

⚡ AI Just Changed Cybersecurity Overnight — Here’s How

Brief Buzz: The UK’s National Cyber Security Centre is sounding the alarm: AI tools are now uncovering decades-old software flaws at a pace the industry’s patching infrastructure can’t match. A massive “patch wave” is incoming — and most organizations aren’t prepared for what’s about to hit.

NCSC’s guidance: prioritize internet-facing systems first, enable auto-updates everywhere possible, and replace legacy tech that can’t be patched.
Anthropic’s Claude Mythos Preview uncovered 2,000+ unknown flaws — including a 27-year-old OpenBSD bug and a 17-year-old FreeBSD remote code execution flaw. Over 99% remain unpatched.
Security firm Theori’s AI scanned Linux’s cryptographic code in about an hour, exposing “Copy Fail” — a 732-byte script granting full root access on every major Linux distribution shipped since 2017.
The gap between “flaw discovered” and “flaw exploited” is shrinking from weeks to hours.

💡 Why Should You Care?

You don’t run servers or manage patches — but your paycheck, medical records, and tax data all live on systems that do. The internet was built assuming bugs get found slowly, by humans, one at a time. That assumption just broke. The vulnerabilities are out there, the clock is running, and the only real question is who finds them next.

🚀 AI Just Beat Harvard’s ER Doctors at Their Own Game

Brief Buzz: A new Harvard-led study published in Science found that OpenAI’s o1 reasoning model outperformed seasoned emergency room physicians at diagnosing real patients — sparking fresh debate about how AI could reshape life-or-death moments in hospitals worldwide.

Tested on 76 real ER cases from Beth Israel Deaconess Medical Center, o1 delivered the correct or near-correct diagnosis 67% of the time at triage, compared with 50–55% for two attending physicians.
With richer patient data, AI accuracy climbed to 82%, while doctors landed between 70–79% — and the AI’s edge was sharpest when information was scarce.
The model especially shone on rare diseases and complex cases drawn from notoriously tricky New England Journal of Medicine scenarios.
Researchers stress it was text-only: no scans, no bedside cues, no real patients affected — and they’re calling for prospective clinical trials before any rollout.

💡 Why Should You Care?

Faster, sharper diagnoses could mean shorter ER waits, fewer missed conditions, and lifelines for under-resourced hospitals. But co-author Adam Rodman warns there’s “no formal framework” for AI accountability yet — and most patients still want a human guiding them through life-or-death calls.

🧠 Inside AlphaEvolve: Google’s Bold New Leap in AI

Brief Buzz: A year after its debut, Google DeepMind’s AlphaEvolve — a Gemini-powered coding agent — is now optimizing algorithms across genomics, quantum physics, logistics, and chip design, delivering measurable breakthroughs from Google’s own data centers to drug-discovery labs worldwide.

Cut DNA sequencing errors by 30% by improving DeepConsensus, helping researchers spot disease-causing mutations more cheaply and accurately.
Designed a counterintuitive circuit now baked into Google’s next-generation TPUs, and trimmed Spanner’s write amplification by 20%.
Drove commercial wins: Klarna doubled model training speed, FM Logistic lifted routing efficiency by 10.4%, and Schrödinger sped up molecular simulations roughly 4x.
Partnered with mathematicians like Terence Tao to crack Erdős problems and break records on the Traveling Salesman Problem.
Slashed quantum circuit errors by 10x on Google’s Willow processor.

💡 Why Should You Care?

AlphaEvolve proves AI isn’t just chatting — it’s quietly rewriting the algorithms behind medicines, power grids, and deliveries. Faster drug discovery and greener logistics could touch your life sooner than expected.

🚀 The AI Boom Is Real — But Most Firms Are Missing the Payoff!

Brief Buzz: Two years into the AI boom and most companies still can’t justify the spend. McKinsey’s senior partner Alexis Krivkovich calls it the “great AI paradox” — the models work fine, but firms are bolting new tech onto old, human-centric workflows.

80% of firms invest heavily in AI, yet few can show their boards real ROI.
At the American Arbitration Association, AI agents now build timelines, analyze cases, and draft summaries — freeing humans to focus entirely on judgment.
The billion-dollar catch: cutting junior hires today means poaching senior talent at a premium in five years, since juniors learn judgment by doing the grunt work.
Building an agentic organization isn’t just tech — it requires rethinking roles, retraining leaders, and making change management the permanent status quo.

💡 Why Should You Care?

If you’re early in your career, AI may quietly close the doors that build expertise. The grunt work that trains tomorrow’s leaders is exactly what’s being automated — risking a serious senior talent crunch ahead.