The tech landscape is shifting fast—AI breakthroughs and Web3 milestones are reshaping how we work, build, and invest. This week’s updates spotlight GPT-5’s rocky launch, reinforcement learning in production, ETH’s rally, and a simple but powerful technique for reducing AI hallucinations.
AI News
GPT-5 Launch: Early Turbulence
OpenAI has released GPT-5, touting deeper reasoning and smarter task routing. But early adopters report glitches, inconsistent quality, and higher costs. While enterprises are upgrading for coding and research, casual users remain divided.
Microsoft Wires GPT-5 into Copilot
Microsoft has integrated GPT-5 across its full Copilot stack—Microsoft 365, GitHub Copilot, and Azure AI Foundry. The update introduces “Smart Mode,” which auto-routes tasks with larger context windows, focusing on workflow grounding over casual chat.
Reinforcement Learning Enhances Ad Copy
Meta reports that reinforcement learning applied to ad text generation improved outcomes compared to supervised models. This shift signals RL and bandit optimization moving from research into large-scale production.
Brains and LLMs: Semantic Overlap
A study in Nature Machine Intelligence reveals that human brain activity during language tasks aligns with LLM embeddings, suggesting these models capture human-like semantics—a breakthrough for neuro-aligned AI interfaces.
DeepMind’s Genie 3: Playable Worlds from Pixels
Genie 3 can now transform video frames into interactive environments, enabling multitask reinforcement learning and sim-to-real testing without manually built levels.
Google’s BigSleep Framework Tested
Security researchers discovered multiple vulnerabilities in Google’s BigSleep AI isolation framework, proving that adversarial testing will continue to expose weaknesses in safety layers.
Web3 News
Liquid Staking Hits Record TVL
The sector has topped $86B, with Lido dominating nearly half the market. Regulatory clarity from the SEC is encouraging institutional participation.
ETH Approaches All-Time High
Rising on-chain activity and ETF inflows are pushing ETH closer to its previous peak, fueling momentum in staking and Layer-2 adoption.
Google Clarifies Self-Custody Wallet Policy
After community backlash, Google confirmed non-custodial wallets are not subject to compliance restrictions—providing relief to developers and distributors.
On-Chain Metrics: Broad Market Upswing
Glassnode reports structural improvements across major assets but warns that liquidity shifts near highs could introduce volatility.
The Evidence-Lock Prompt: A Simple Fix for AI Hallucinations
One of the biggest challenges with AI systems is hallucination—when a model confidently generates information that isn’t true. That may be fun in art, but it’s risky in business reports, compliance, or executive updates.
At Tesseract Academy, we’ve been testing practical methods to reduce hallucinations. One of the most effective is what we call the Evidence-Lock Prompt (ELP).
What is the ELP?
The ELP is a structured prompt that forces the AI to only use predefined evidence. If data is missing, the model must reply with “Not in evidence.” No guessing.
This makes outputs:
- Focused → No wandering into outside knowledge
- Trustworthy → Every claim can be traced to data
- Practical → Perfect for reports, audits, and compliance
Plug-and-Play Template
Use only the facts inside the EVIDENCE box.
If any required info is missing, write “Not in evidence.”
Do not guess. Write the final answer only.
EVIDENCE:
[Fact 1]
[Fact 2]
TASK: [What to produce, who it’s for, length limit, tone]
FORMAT: [Bullets/JSON/Table/etc.]
Example:
EVIDENCE:
- Q3 MRR £412k
- Churn 2.1%
- Top driver: onboarding emails
TASK: Draft a 120-word executive update with one next action.
FORMAT: 3 bullets + 1 action
➡️ The output stays grounded in evidence and suggests a concrete next step.
Next Steps
At Tesseract Academy, we specialize in helping professionals and teams harness AI effectively. If you’d like more ready-to-use prompts, frameworks, and practical AI training:
👉 Explore our resources here
💬 And if you’d like bite-sized AI tips & templates, join our WhatsApp group: Join here